Está en la página 1de 421

Foreword Series

Computational neuroscience is an approach to understanding the information content of neural signals by modeling the nervous system at many different structural scales , including the biophysical , the circuit , and the systems levels. Computer simulations of neurons and neural networks are complementary to traditional techniques in neuroscience. This book series welcomes contributions that link theoretical studies with experimental approach es to understanding information processing in the nervous system . Areas and topics of particular interest include biophysical mechanisms for computation in neurons , computer simulations of neural circuits , models of learning , representation of sensory information in neural networks , systems models of sensory -motor integration , and computational analysis of problems in biological sensing, motor control , and perception . Terrence J. Sejnowski Tomaso A . Poggio

Preface

Neural networks began to be the object of serious researchin the 1940s . They did not becomewidely known or popular until the 1980swhen several academic centers published articles and distributed software that enabled individuals with modest mathematicaland computing skills to learn computational network principles and apply them to broad rangesof projects inside . At 6rst, neural networks university settings as well as in extramuraldomains were predominantly applied to industrial problems in areassuch as pattern . However, many active neural network recognition and signal processing researchers were guided by insights from neuroscience and psychology, and believed that neural networks could become an important computational es. techniquefor furthering our understandingof brain and cognitive process This belief was borne out by signi6cant progress in the 6eld during the 1990s when substantial improvements in computer architecture , software , and computing speed occurred . In particular, interest in applying design neural networks to understandingneuropsychology and mental function and dysfunction has increaseddramatically. Other colleaguessuch as neurolo-disciplinary neuroscientistsalso have facilitated , and cross gists, psychiatrists interest in this new 6eld. The feasibility of making connections between models and clinical data has also increased , calling for a textbook that brings much of this together growing body of work in a single volume. This book is the 6rst of its kind that is speci6callyintended for neuropsychologistsand related disciplines . Our book has models of many widely used neuropsychologicaltests and tasks , including the Wisconsin Card Sorting, Stroop, verbal fluency, Tower of Hanoi, Line Cancellation , and many other tests. It covers a wide range of ' ' , including Alzheimer s disease syndromes , Parkinsons disease , schizophrenia , alcoholism stroke , attention , , epilepsy de6citjhyperactivity disorder, and frontal lobe disorders . The modeling techniquesutilized are not dominated " one school " of modeling; rather, they include widely used modeling by , such as backpropagation and adaptive resonance paradigms , and models for a to brain systems approach interacting designed regions implicated in . In addition, some chaptershave been included to facilitate a cognitive tasks

better understanding of the neurobiological .and neuroanatomicalbasis of cognitive network models. Prerequisitesof a technical nature are minimized in this book. The mathematics involved in the network models are deemphasized , and in most cases the theoretical basisfor the model' s structure can be discernedfrom the dia, the aim is to present neural network grams combined with the text. Hence as theoretical tools that can be readily learned and applied by techniques and other neuroscientists , psy(including neurologists neuropsychologists chiatrists , clinical psychologists , mathematicians , and computer scientists ) with particular interest in clinical patient applications . We hope that this will contribute to the accessibilityof neural networks and to an intensification of the ongoing dialogue between the clinical and theoretical communities .

preface

Contributors

J. WessonAshford, MiD ., Ph.D. Department of Psychiatry and Neurology University of Kentucky College of Medicine Lexington, Kentucky , Ph.D. RajendraD. Badgaiyan of Department Psychology University of Oregon , Oregon Eugene ., PhiD. , M.D JeanP. Banquet CREARE and , Neuroscience Modeling Institute of Neurosciences Pierre and Marie Curie University Paris , France Yves Burnod, PhiD. INSERM-CREARE , Neuroscience and Modeling Institute of Neurosciences Pierre and Marie Curie University Paris , France Nelson Butters , PhiD. (deceased ) of Neurosciences Department University of California, SanDiego SanDiego, California

, Ph.D. Agnes S. Chan of Department Neurosciences University of California, SanDiego SanDiego, California and Departrnmentof Psychology, ChineseUniversity of Hong Kong Shatin , New Territories Hong Kong -Pierre Changeux , PhiD. Jean Molecular Neurobiology PasteurInstitute Paris , France Kerry L. Coburn, Ph.D. Department of Psychiatry Mercer University School of Medicine Macon, Georgia JonathanD. Cohen, M . D., PhiD. Department of Psychology CarnegieMellon University and Department of Psychiatry University of Pittsburgh Pittsburgh, Pennsylvania Laurent Cohen . , MiD ., Ph. D Department of Neurology SalpetriereHospital Paris , France

, MiA . John Cardoso of JoseL. Contreras-Vidal, PhiD. Department Psychology of New Brunswick University Department of ExerciseScience Frederic ton , New Brunswick , Canada Arizona State University , Arizona Tempe

Antonio R. DamasioMiD ., PhiD. Department of Neurology University of Iowa College of Medicine Iowa City , Iowa Hanna Damasio , MiD . Department of Neurology University of Iowa College of Medicine Iowa City , Iowa StanislasDehaene , PhiD. -CEA INSERM Unite, SHFJ Orsay, France Martha J. Farah . , Ph. D of Department Psychology University of Pennsylvania , Pennsylvania Philadelphia ., PhiD. ,M.D JoaquinM . Fuster of Psychiatry Department UCLA School of Medicine Los Angeles, California , PhiD. Philippe Gaussier ETIS ENSEA University of Cergy-Pontoise , France Cergy-Pontoise , MiS. Angelika Gissler Neuroscience and CREARE , Modeling Institute of Neurosciences Pierre and Marie Curie University Paris , France Dylan G. Harwood, M .A . Department of Educationaland PsychologicalStudies University of Miami Miami , Florida Michael E. Hasselmo , D .Phil. Department of Psychology and Program in Neuroscience Harvard University , Massachusetts Cambridge

] . Allan Hobson, MiD . Laboratory of Neurophysiology Department of Psychiatry Mental Health Center Massachusetts Harvard Medical School Boston, Massachusetts SamLeven , PhiD. ScientificCybernetics , Maryland Chery Chase DanielS. Levine, Ph.D. Department of Psychology University of Texas at Arlington Arlington , Texas Debra L. Long , Ph.D. of Department Psychology University of California, Davis Davis, California . Roderick K. Mahurin, Ph. D Battelle Research Center and Department of Psychiatry University of Washington Seattle , Washington Raymond L. Ownby , M . D., PhiD. Department of Psychiatry University of Miami School of Medicine Miami , Florida , Ph. D., Psy. D. Randolph W. Parks , Program Neuropsychology of Department Psychiatry SouthernIllinois University School of Medicine Springfield, Illinois . Michael I. Posner , Ph. D of Department Psychology University of Oregon , Oregon Eugene David P. Salmon , Ph. D. Department of Neurosciences University of California, SanDiego SanDiego, California

Discussions and Debates

-Schreiber . David Servan , MiD ., Ph. D of Psychiatry Department University of Pittsburgh Medical Center ShadysideHospital Pittsburgh, Pennsylvania Chantal E. Stem, D .Phil. Department of Radiology Massachusetts GeneralHospital Harvard University Charles town , Massachusetts ., PhiD. Jeffrey P. Sutton, M . D Neural SystemsGroup, Department of Psychiatry GeneralHospital Massachusetts Harvard Medical School town , Massachusetts Charles . Lynette J. Tippett , Ph. D of Department Psychology The University of Auckland Auckland, New Zealand Daniel Tranel, PhiD. Department of Neurology University of Iowa College of Medicine Iowa City , Iowa Bradley P. Wyble, M . A. Department of Psychology and Program in Neuroscience Harvard University , Massachusetts Cambridge

Contributors

An Introdudion to Neural Network Modeling : Merits , Limitations , and Controversies

L. Long Debra W. Parks , Randolph , and . Levine DanielS

Many researchersin cognitive psychology and neuropsychology have describedhuman mental activity in terms of abstract information-processing models . Still other researchers have focused on regional brain areas associated with hypothesized neuropsychological test performance without an ' appreciationof multiple brain regions contributions to cognitive test performance . However, thesetraditions have been supplementedin recent yearsby efforts to develop models of cognitive process es that are better founded in our understandingof neural circuitry . Researchers who have adopted a neural network approach attempt to model cognition as patterns of activity acrosssimple computing elementsor units (Feldman& Ballard , 1982; Gross 1982 1988 McClelland Rumelhart & the POP Research , ; , , berg , Group, 1986; Rumelhart& McClelland, 1986). Such models have been used to investigate theoretical claims about the nature of various cognitive process es (e.g ., pattern attention semantic natural , , , recognition ) memory language processing and to solve a wide range of applied problems (e.g., machinevision, medical , speechrecognition and production). diagnosis In this chapter , we use the term neuralnetworkbroadly to refer to a large class of models that share certain architectural and processing features : (1) thesemodels contain many simple processingelementsor units that operate in parallel ; (2) the units in these models communicateactivation along connections (weights) of varying strength; and (3) the models learn by adjusting their weights as they gain experiencein some environment. Models of this models and paralleldistributedprocessing type have also been called connectionist models . Neural network models were introduced to mainstreampsychologists and , including Feldmanand computer scientistsin severalinfluential publications Ballard (1982), Hopfield (1982), and Rumelhartand McClelland (1986). Since that time, thes .e models have been the topic of substantial theoretical and computationalinterest and the sourceof considerablecontroversy. The main ideas involved in neural network modeling have a substantial , 1988; Levine, 1983, 1991; and Parks et history (seeAnderson & Rosenfeld al., 1991 for accounts . An influential article in the field was published , ) partial by McCulloch and Pitts in 1943. They demonstratedthat any logical

function could be duplicated by a suitable network of all-or-none neurons with thresholds . Biological underpinnings of neural network theory were addedby two researchers in particular: Hebb (1949), who formally proposed that learning dependson synaptic weight changesin responseto paired preand postsynaptic stimuli; and Rashevsky (1960), who proposed methods for averaging the all-or-none behavior of single neurons into continuous . Rosenblatt (1962), Werbos (1974), and dynamics of large neural ensembles others combined McCulloch and Pitts' s insights with Hebb' s in developing the ancestorsof current error-correcting learning networks, such as Rumel ' hart and McClelland s (1986) backpropagation network. Grossberg (e.g ., 1969) pioneeredthe developmentof continuous dynamical system networks that embodied psychological principles such as associative learning and lateral inhibition. Such principles were also explored by several other researchers who are still leaders in the neural network field, such as Amari Anderson (1971), (1968, 1970), and Kohonen (1977). Our purpose in this book is to make the neural network approach accessible to practicing neuropsychologists , clinical psychologists , neurologists , and research neuroscientists mathematicians and , ( including psychiatrists ) who may have little substantiveknowledge of the topic. computer scientists The chapters gathered here describe recent advancesin the application of neural network modeling to a wide range of topics. Part I provides an overview of neural network modeling. The chaptersin part I describe the basic architecture of neural networks and discuss some of the assumptionsthat underlie this approach . In addition, the chaptersdescribe the application of these models to theoretical topics such as attention and hippocampal contributions to memory. The chapters in part II describe neural network models of behavioral states , such as alcohol dependence , learned helpless ness . Part III is devoted to neural network , depression , and waking and sleeping modelsof neuropsychologicaltests suchas the Stroop, Tower of Hanoi, and Line Cancellationtests. In addition, the chaptersin this section describe -specific topics such as schizophrenia the application of these models to syndrome , acalculia , neglect , attention deficit/ hyperactivity disorder, and lexical retrieval. The final section , part IV , describesthe application of neural network models to dementia . Thesechaptersdescribemodels of acetycholine ' s disease and Alzheimer' s and memory, frontal lobe s~ drome, Parkinson , disease . The Wisconsin Card Sorting and verbal fluency tests are modeled within the dementiaframework. Our goal in this first chapter is to provide a comprehensiveintroduction to principles of neural network modeling. In the first section , we describethe basicarchitectureof neural networks and contrast thesemodels with the traditional symbolic models used in artificial intelligence (AI ). In addition, we , provide details about two of the most popular classesof network models networks and networks based on resonance theory backpropagation adaptive , we discuss some of the controversies surrounding (ART ). In the second section the neural network approach . Should cognition be modeled using

Introduction to Neural Networks

principles of language and logical reasoning, or principles of the underlying neural architecture ? Should neural network models be viewed as simulations of cognitive theories? Can behaviors that appear functionally modular (e.g ., visual vs. semantic processing, lexical vs. syntactic processing ) be performed by a system that is not modular in the same way ? In the final section , we describe some of the theoretical and computational limitations of current neural network models .

-PROCFSSING AN AL TERNA nVE TOTRADITIONAL SYMBOL


MODELS
Traditionally, cognitive psychologists have formulated their theories by , 1980, 1982; Newell, analogy to computersand computer programs ( Newell Rosenbloom , & Laird, 1989; Pylyshyn, 1989). This analogy has led to a symbolic- or information-processingview of human cognition. In the symbolic , goals , beliefs , and knowledge are representedas abstract approach . These symbols ; that is, they symbols are semanticallyinterpretable elements can be describedin terms of ordinary conceptsrelevant to the domain. Intellectual tasksare performed in these systemsby meansof explicit rules, such as the rules of formal or sentential logic (Fodor , 1976; Fodor & Pylyshyn, 1988; Newellet al., 1989). These rules specify how symbols are to be . In this approach , symbols and the rules that manipulatedand concatenated on them are in discussed the absence of assumptionsabout brain operate location or brain process es. " " In contrast , the neural network approach is an attempt at brain -style computation. Here are a few of the properties that distinguish neural networks from traditional symbolic models : 1. Neural networks incorporate the basic properties of neural circuitry . Processing takes place in a system of connectedmodules . Eachmodule consists of a population of interconnectedcomputational units. Eachunit is a simple information-processingdevice that (a) computesa numericalactivation value from the input of other units and (b) communicatesthis activation value to other units along connections of varying strengths (Feldman & Ballard , 1982; Feldman , Ballard , Brown, et al., 1985; Rumelhart& McClelland, 1986). A unit may correspondto a single neuron, a population of neurons , or a feature of an object or concept . Someunits receive input from outside the system . Several ; other units produce output or responsesthat exit the system classes of neural network models contain "hidden " units. Hidden units have connectionsto other units in the system , but do not send or receive information from outside the system . 2. The individual units in neural networks perform computations in parallel. The activation of each unit changesconstantly in responseto the activity of the other units to which it is connected . The massively parallel nature of neural networks correspondsto an important feature of brain organization

: Levine : Neural NetworkModeling ,& Long , Parks

and processing (Anderson & : Hinton , 1989; Pellionisz & : LIinas , 1982; Sejnowski, 1986). Although an individual biological neuron is a slow processing device, large numbers of neurons operating in parallel can accomplisha task very quickly (Feldman& : Ballard , 1982). 3. Neural networks encode knowledge very differently than do traditional : Ballard : McClelland, 1986; , 1982; Rumelhart & symbolic models (Feldman& . In neural Smolen 1988 1989 networks , , , ) sky knowledge is encoded in the of the connections between individual strengths processing units. Representations consist of patterns of activation acrosslarge populations of conneded units. An individual unit may participate in several distinct . representations 4. Neural network models also perform mental computations differently than do symbolic models . The rules in a neural network operate at the level of the individual units (Rumelhart & : McClelland, 1986; Smolensky , 1988, 1989). These rules specify how units compute their activation value, how activation is passedto other units, and how connection strengths are modified. Intelligent behavior arisesfrom the way in which interacting units are . connected 5. Neural networks learn from experience . The models discover a set of connection strengths (weights) that capture the internal structure of a domain. These connection strengths are typically found by meansof errordriven adjustmentsin the connectionsamong units (Grossberg , 1976; Hebb, 1949; Hinton & : Sejnowski 1986 Rolls & : Treves Rumelhart 1998 , ; , , Hinton , ; & : Williams, 1986; Widrow & : Hoff , 1960). Once a neural network has discovered a set of weights that results in appropriate behavior, the model can repeata pattern of activity at somelater time or produce a pattern of activity when given some portion of it as a cue. Thus, these systems " program " themselves to perform extraordinarily complex tasks . These features of the neural network architecture give rise to properties that make neural networks an attractive alternative to traditional symbol. First, neural networks have distributed representations ; processing models thus, they seem well suited to model process es that appear to operate on -like representations , such as the mental rotation of objects (Cooper, analog : Cooper, 1982; Shepard& : Metzler, 1971). In addition, neural 1976; Shepard& networks have knowledge that is distributed in the connectionsbetween units. Thus, these networks have little difficulty modeling nonverbal or intuitive process es. Second , neural networks provide good accounts of pattern recognition, motor control, and associativelearning. Such tasks appearto require mechanisms for processingmassiveamountsof information in parallel while satisfying . Neural networks are appropriate large sets of probabilistic constraints " " for these tasksbecause soft constraints (Smolensky , 1988, they implement 1989). The connectionbetween two units is a constraint in that it determines whether or not one unit should be active when the other is active. This con-

Introduction to Neural Networks

" " straint is soft because it can be easily overridden by the input of other units . The network behaves in a way that balances a massive number of " these soft constraints . Thus , the network settles" into stable representational patterns . As a consequence, neural network models tend to be less brittle and rigid than traditional symbolic models . Third , neural networks are attractive alternatives to traditional AI models because they exhibit spontaneous generalization . These models can respond appropriately to novel stimuli once they have learned the internal structure of a domain . Similar representations typically have many units in common . ' Thus , a new item can be classified in the network s existing representational scheme when its features activate representations that share similar features. This is an advantage for tasks that require generalizations &om large sets of stored exemplars or tasks that involve degraded input . " Finally , neural networks are attractive because they exhibit graceful " degradation . A representation consists of a pattern of activation that is distributed over a population of units . The pattern of activation representing an entity is distinctive ; therefore , different entities can be represented using the same set of units . This feature of neural networks makes them resistant to damage. A distinct pattern of activation may still occur even if the system has lost some portion of its units . In addition , graceful degradation allows for relearning after damage. A damaged network may retain a subset of the weights that produced appropriate behavior before the damage . To the extent that these weights capture some of the internal structure of the domain , the network should readily relearn associations that it previously knew . So far , we have discussed some of the characteristics that are common to a variety of neural networks . Next , we describe in some detail two of the most popular classes of neural network models .

Networks1 Backpropagation
Since the middle of the 1980s, backpropagation networks became popular with the publication of a collection of articles from diverse fields and accompanying software ( McClelland et al., 1986 ). The simplified , though powerful , software release enabled university neuroscientists and other researchers with moderate computer skills to program rudimentary cognitive processes with relatively little assistance. Previously , these programming techniques were largely confined to supercomputer centers, such as the large government facility at Los Alamos . In this section , we examine some of the tenets of backpropagation by first describing the basic components of a neural network at the cellular level . This is followed by a brief discussion of how the terms parallel and distributed are essential to this process. We then describe some of the important neural network principles of backpropagation in training a neural network . We illustrate these principles using a standard neuropsychological test of executive functioning in humans, Tower of Hanoi . A more detailed in vivo demonstration of backpropagation with the

t Levine : Neural NetworkModeling ,& Long , Parks

Tower of Hanoi in a clinical population of left frontal lobe patients can be found in chapter 9 by Cardoso and Parks. Ownby utilizes backpropagation in chapters on alcohol dependence (chapter 5 ), cortical neglect (chapter 11) and attention deficit / hyperactivity disorder (chapter 11). Servan- Schreiber and Cohen in chapter 8 describe the use of backpropagation to model the Stroop test in schizophrenia . Finally , a hypothetical model of basal ganglia function in chapter 14 by Mahurin is applied to human movement sensori ' motor dysfunction (e.g ., Parkinson s disease).

Artificial Neural Networks : Computer Architecture at the Cellular Level In order to understand some of the essentialsof backpropagation , the readerfirst needsto appreciatehow a computer neuroscientisttranslates principles of dynamic living tissue into a languagethat the computer understands . Historically, we draw on such pioneers such as Hebb (1949) and his ' Learning Law, and McCulloch and Pitts s (1943) M -P neuron. Another pioneer was Rosenblatt(1962) who invented the Percept ron, a simple computa tional learning device. He wrote, " Perceptronsare not intended to serve as detailed copies of an actual nervous system . They are simplified networks, to the of lawful study designed permit relationships between the organization of a nerve net, the organization of its environment, and the psychological of which it is capable ." performances Currently, one may describea neural net as made up of very simple com, neurodes , processing putational units. These units are also called neurons elements , cells, nodes , and various other namesdepending on the theoretical of the neuroscientist . The seminal conceptsbehind the creation preferences of artificial neural nets, along with the conceptualfoundations of the computational scienceswhich deal with the properties of these units, give rise to the current essentialfeaturesof a computational neuron. Thus, the artificial neuron is a crude eclectic simulation of its biological counterpart. By adapting , the artificial neuron is said to consist languagefrom the biological sciences of its own dendrites . It receives , synapses , cell body , and axon terminals stimulation from nearby cells , or from its environment, and generates a modified action potential or nerve signal. A frequently used model of a neuron is shown in figure 1.1. This model neuron (Parks & : Cardoso , 1997) has three inputs: Xl , Xl , and . Theseinputs potentially receivetheir stimulation from the output (axons) X3 of other neurons , or directly from sensorycells monitoring the environment. stimulation ) reach the body of the neuron through coupling Input signals ( . The efficacy of these synapses , known as weights of the connections synapses . The signals , are representedhere respectively as WI, Wl, and W3 from inputs Xl and X3 reach the neuron through inhibitory synapses(synapses with negative weights). The signal from input Xl reachesthe neuron through an excitatory synapse(synapsewith a positive weight). Before the , they are scaled by the weights of their input signals reach the neuron and added together so that the net-value of respective coupling synapses

Introduction to Neural Networks

Figure 1.1 A simple model of an artificial neuron.

= xi 1 = w1 5 U ) = 3 = + 1 x2 w y " 5 _ . 6 ' . . ~ = 1 x3
output
stimulation I (or weighted sum) reaching the neuron will be, assumingthat eachinput has one unit of stimulation, I = Xl * WI + Xl * Wl + X3* W3= 1 * ( - 5) + 1 * 10 + 1 * ( - 2) = 3 Sincethe input net-value I, in this simple case , is transferreddirectly without loss of transformation to the experiencingany output of the neuron, its is also to 3. This is an overly simpli6ed model, but it may outputy equal give readersunfamiliar with this material an idea of how a biological neuron can be simulated . Ordinarily , in a working model, there are other parameters such as a bias , an input threshold , an input (fan-in) function that further an the activation function that determinesthe activation , reshapes input signal (or excitation) of the neuron, and an output (fan-out) function that changesits activation value to the Anal value that will eventually reach the postsynapticneurons) or the output of the network. An arti6cial neural network is made up of many of these simple units, highly interconnectedand generally organizedby layers with eachneuron in a given layer connected to every neuron in the succeedinglayer. Signals travel through these connections from unit to unit and Anally generate an output. Some nets are capable of learning by themselves (unsupervised learning). The neuronsin thesetypes of nets continually monitor their inputs and keep adjusting their weights trying to mirror the signalsthey receive .A -learning net is the Kohonen network well-known type of unsupervised , 1984). These types of networks (self-organizing) are capable of (Kohonen modeling the probability distribution function of the input data. Other nets need an executor (supervisedlearning) to tell them whether they are correct in their responses , and, if not, by how much they are wrong . The difference between the expected(correct) responseand the actual responsegiven by a unit is used , to readjust the weights of the connectionsand further , as feedback to reduce the error until the unit eventually generates the correct . This is a gradual processwhich usually takes many iterations to response . accomplish After a neural net is trained, we say that the net has learned and that the has been stored (as values ) in the weights of its connections acquired knowledge . For example , if the correct output that we expectedfrom the neuron in 6gure 1.1 was a value other than 3, we could teachthe neuron by

: Levine: Neural Network Modeling ,& Long , Parks

, and W3 so that the unit eventually gradually adjusting the values of WI, W2 generatesthe desired output. Two important fundamental features of network , the parallelism of neural nets, and the distribution and organization in thesenetworks, are describedin more detail in the of information storage next section . Parallel Distributed Processing The term parallel refers to the type of network architecture (Koch & : Segev , 1998; Parks et at., 1991). It identifies the way the neural net is organized and process es information. In a parallel , information is not processedsequentially , one step at a time, or by system one unit at a time, but simultaneouslyby many units. This is an important feature of artificial neural networks becauseit seemsalso to be a feature of neural systems in the human brain. The computational architecture of the brain appearsto be highly parallel. Biological neurons are slow devices and the traveling speeds of nerve signals are slow. The brain is capable of achieving its rapid processing speed only through the massive parallelism of its neurons . Similarly to the brain, when an input signal reachesa neural net, the signal is simultaneously distributed throughout many neurodes . Depending on the type of network, one whole layer or even the whole network es the signal may be involved at the sametime and every unit process . The in and , efficiency , concurrently gains speed computational power are enormous . However, all of these gains substantially increasethe complexity of the process . The term distributedrefers to the way in which information is stored and , whether natural internally in the network. An intelligent system represented or artificial, can only learn , remember , and recognize if it is capableof keeping some internal representation of what it has learned . AI has utilized rulebased methods and symbolic programming techniquesas its main tools to represent and manipulate knowledge in a computer. Semantic networks : Abelson, 1977), , scripts , and frames (Schank & (Quillian, 1968), schemas among others, are powerful knowledge tools, but they have not proved to be sufficient to advancesubstantially the study of learning . Connectionism offers a very different approach . In artificial neural networks, learning and are hard to differentiate as two independentproknowledge representations cesses . It is generally accepted that, in neural nets, it is the weights (the . The weights encode ) that do the representing strength of the connections whatever knowledge the network has acquiredduring its learning phase . Two main types of representationare possible in neural nets: local and distributed (Rumelhart & : McClelland, 1986). In local representations , individual units stand for someunique concept or feature of the input; a particular featureof the input causes one single unit to becomeactive. In distributed one , representations concept or feature of the input is representedby the activation of a set of units. In other words, a single input causesmany hid~ den units to becomeactive (Hanson & : Burr, 1990). The multilayer neural net

Introduction to NeuralNetworks

in clinical . Figure 1.2 TheTowerof Hanoitestasused by neuropsychologists practice that was used in the current experiment contained distributed knowledge representations . Many units become active for a given input . However , we have not strictly adhered to how the distributed knowledge representations are programmed nor to how they are interpreted . For example , we do ascribe some computer architecture elements to frontal functions , attentional factors , and visual processing . A more detailed discussion of the controversies of localization and modularity of brain functions is addressed later in this chapter .

Backpropagation with a Complex Human Neuropsychological Test In this section , we define someof the featuresneededto model a well-validated neuropsychological test, Tower of Hanoi (TOH ). In its classic form, the TOH puzzle consistsof three vertical pegs mounted on a board and a set of disks or rings graded in size (Cardoso , 1991; Anderson & Kushmerick , 1992; Shall ice, 1982). Initially , all of the rings are stackedin order of size on one peg, with the largest ring at the bottom and the smallest at the top (figure 1.2). The TOH requires the ability to anticipate subroutine steps to achieve the desired final position of the rings. The backpropagation network has three layers (figure 1.3): input, hidden, and output . The input layer has , the hidden layer has fifteen, and the output layer twenty -seven neurodes has twenty -seven . The sizesof the input and output layers were selectedto match the number of featuresthat we wanted to represent . The size of the hidden layer was based on the number of statistically significant factors in the input data. A more comprehensivedescription of subroutine and programming is provided in chapter 9. parameters In our training of this feedforward multilayer network, the back BP propagation ( ) program utilizes an algorithm that is basedon the generalized delta rule (Caudill & Butler, 1990; Chauvin & Rumelhart , 1995; Pao , 1989; Rolls & Treves , 1998). Briefly, without discussing the mathematical . In the first step, the principles that underlie the rule, BP requires two steps is and input presented propagated forward through the net to activate the output units. The output value of each of these units is then comparedwith

: Levine : Neural NetworkModeling .& Long , Parks

unit 1 2 unit 3 27 unit


Input Layer Output Layer unit 1 unit 2

unit3

unit27

Figure 1.3 Architecture of the Tower of Hanoi network used to study the results of degrada . tion in the output-layer connections

. The second step is the expected (target) value and an error is is computed a backward pass through the net to compute the is for each unit in the , we compute the weight error network. Once these two stepsare completed . Users of this derivatives, and then compute the required weight changes . descent is a BP the that aware be rule must procedure gradient algorithm the function error Gradient descent proceduresminimize the by following contour of the error surface , always moving downhill and settling at the bottom. In complex error surfaces , the processmay settle into a local minimum not the lowest point on the surface but ) rather a low is which ( point , ' s lowest . Gradient surface the minimum than into the desired global ( point ) . One of the parameters descentmethods require very small steps to succeed used in training the network is the learning rate, 8, a constant of proportionality that determines the magnitude of the weight changesat every step. . Large steps can causeoscillations during Large 8-values mean larger steps learning. In other words, one can easily visualize the system trying to settle into a valley, but always bouncing from hill to hill . Choosing the largest possible value that does not causeoscillations leads to the fastest learning , 1989). , 1995; Pao (Chauvin & Rumelhart . This parameter has the Another important BP parameter is momentum effect of a high-frequencyfilter. Filtering out the high-frequencyvariations in , and allows for the error surfacepermits the use of relatively large 8-values a distributed brain As in the . faster learning , representationprotects the system from forgetting all that it has learned in the caseof injury or mutilation ' , the systems performancesuffers ). In such cases (in programming language . This little or no degradation , and the system is not renderedtotally useless to the allows descent system processvery large data methodology gradient sets , as well as detect very subtle changesduring cognitive modeling. This methodology may potentially clarify the multiple system neuropathological conditions, such as the gradual cognitive changesof insidious neuropsychiatric ' . decline seenin Alzheimer s disease

o Neural Networks In~ roduction ~

INPUT
networkcombining associative . Figure 1.4 Generic categorization learningand competition from Levine of Miller Freeman ( Modified , 1989 , with pennission Publishers .) Adaptive

Resonance ) Theory (ART

Backpropagationnetworks have achieved remarkable successnot only in simulating a wide range of neuropsychologicaldata, as will be seenin later chaptersof this book, but in a .wide range of engineering applications that " " . Such networks are fairly homogeneous in require intelligent systems structure and in mathematical properties. They do not , therefore , capture much of the richnessand diversity of the structures of actual brains. Also , " " , in the form of an backpropagationnetworks rely on outside supervision " " input telling the network what the desiredor target responseshould be. There is another long tradition in neural network theory (Levine, 1991) based on networks that are self-organizing and decomposableinto smaller networks that have specializedcognitive properties themselves . Of these , one of the best known is the adaptiveresonance theory (A Rn network for , introduced by Carpenter and Grossberg (1987). ART is classifying patterns built on earlier networks developed by Grossbergand his colleagues(Gross, 1982) that embody principles such as competition and associative berg, 1976 learning. A generic categorization network, examplesof which appear in von der Malsburg (1973) and Grossberg(1982), is shown in figure 1.4 (Levine, 1989). The nodesat the Fl level representsensoryfeatures ; the nodesat the F2 level . .representcategories When an input arrives from the outside environment, it activates the corresponding nodes at the F1 level, which thereupon sends

: Levine : Neural NetworkModelin8 ,& Long , Parks

of

F2

jth

sum

WiJ

to

Is

Xi

Xi

Signal

node

8 8 8 t : : : ' : : : ~ 8

8 ~ t E Wij8 t

8 8 8

F2 (CATEGORIES )

F1 (FEATURES )

ST N RE , + WA + Xi . , . ... GAIN CON F1 + + +


ATTENTIONAL SUBSYSTEM ORIENTING SUBSYSTEM
INPUT PATTERN asavector . Sensory ) network theory(ART input to FI is stored Figure 1.5 Adaptiveresonance " " . The FI node activitiesare multipliedby the bottom-up connection of node activities . Then the largest to F1 , thentransmitted ., wherethe onenodereceiving signalwins weightsWjj " " -down 's stored the as with that node , the input is compared by top prototype represented to the finetuning device . Gaincontrolis a mathematical , not essential connection weightsWjj to inputsthat are is a response . Reset to novelinputs . Orientingis a response herein discussion and Grossberg , with a prototypeand mismatmit. (AdaptedHornCarpenter , 1987 compared Press .) of Academic with permission

bottom -up signals (weighted by the connedion strengths ) to the F2 level. Whichever F2 node receives the largest signal determines the category . The signal to the appropriate category in which the input is classified node inhibits activities of the other nodes (as indicated by the minus sign in figure 1.4). However, Grossberg (1976) showed mathematically that a categorization network such as the one in figure 1.4 can lead to instabilities in coding if the . He showed that such same set of input patterns is presented repeatedly -up signalsfrom Fl bottom the be can instability preventedby supplementing to to F2 with top-down feedback signals from F2 Fl, leading to the ART model shown in figure 1.5. These top down signals allow for comparisonof : Keele , the input with previously learned category prototypes (seePosner & stored at data relevant for 1981 : Rosch ) 1970; or Mervis & , , psychological with the . If an input is closeenough to a prototype , it resonates the F2 nodes ) and is stored at the correspondingnode. category (hencethe word resonance The term adaptivemeansthat the prototype is updated by meansof a time' , one s mental prototype weighted averagewith the new input: for example

Introduction to Neural Networks

Yi

~ IIII ~ ~I II~

Wji

of the category "bird " might look like a sparrow until one has seen many robins, and then changesto some sort of weighted averageof a sparrow and a robin. Thus, in the ART model the top- down feedbackconnection weights, as well as the bottom-up connection weights, are changedby paired activities of the two interconnectednodes , which is a variant of a learning principle enunciatedby Hebb (1949). Adaptive resonanceis one of the principles used by Grossberg and his , 1988) in models of a wide range of cognitive phenomena colleagues(Grossberg including classicalconditioning , preattentive vision, word recognition , and many others. It combineswith other principles that are suggested by both physiological and psychological data. One of theseprinciples , which also plays a role in ART , is lateral inhibition as a meansof modeling decisions between activations of competing entities, whether they be percepts , , drives, beliefs , or motor plans. Another principle is opponent categories , which models selective enhancement of novel stimuli or changing processing reinforcement contingencies . More recently, a variety of researchers have applied different combinations of all of these network principles to model aspectsof neuropsychological functioning and mental illness (Hestenes , 1998; Levine, 1996; and chapters4 and 15 in this book).

CURRENT DEB ATFSIN NEURAL NETWORK MODELING


In this section , we discussa few of the controversiessurrounding the neural network approach . These controversies relate to three issues : (1) Does a of ? (2) Are neural theory cognition require symbolic representationsand rules networks simulations of cognitive theories ? (3) Is the functional modularity of many cognitive activities reflectedin the underlying architecture ?

The Needfor Symbol i ( : Representations and Rules


Neural network models cannot be distinguished from traditional symbolic models on the basis of a belief in the existenceof representations(Fodor & , 1989). Representationsexist in both types Pylyshyn, 1988; Smolensky of models , although they are realized somewhat differently. In symbolic models , representationsare semantically interpretable in that they correspond to the mental statesthat we ascribeto persons(e.g ., beliefs , thoughts, desires , etc.). Symbolic representationsare viewed as purely formal objects. This means that symbolic representationscan be manipulated by virtue of their form or syntax without regard to their semantics . In neural network models , representations are not primitive computa tional elements(Smolensky 1988 1989 rather , , ); , representationsare patterns of activation across individual units. Although these patterns also correspond to the mental states that we ascribe to persons , the individual units that participate in a representation mayor may not be semantically . interpretable

: Levine: Neural Network Modeling ,& Long , Parks

The distributed nature of representationin neural network models is the . Some critics have argued that the sourceof one objection to this approach absenceof semanticinterpretations for the units that participate in a representation leadsto models that are as inscrutableas the inner workings of the : Bruce : Pylyshyn, 1988; Massaro brain itself (Burton & , 1988; , 1994; Fodor & the . However of 1990 1991 , , ; , ) McCloskey approach proponents Suppes have argued that individual units do have interpretations in the sensethat : Rumelhart , they representmicrofeaturesof a domain (Hinton , McClelland, & " namable " entities 1986). Although thesemicrofeaturesmay not correspondto , they have meaning in the sensethat they map onto some property of the domain. One of the challengesfaced by neural network modelers is to discover how microfeaturescombine to representsemanticallyinterpretable . concepts The property that truly distinguishes neural networks from symbolic models concerns the manner in which representationsare manipulated and are governed by rules that . In symbolic systems transformed , representations are themselves represented symbolically (Newellet al., 1989; Pylyshyn, " " 1989). These rules take the form of if A , then B. In a typical AI model, . Thus, representations these rules are implemented serially to make inferences interact with (i.e., causally affect) other representationsin a manner specifiedby the rules. The rules in a symbolic system operate on representations in a manner that is sensitive to their constituent structure. That is, can have a complex structure in that they are composed symbolic representations of symbols that stand in syntactic relations to other symbols. Rules can operate on a symbolic representation by virtue of its structure alone : Pylyshyn, 1988). (Fodor & In neural network models , the relation between representationsand rules are not viewed as formal objects is very different. Representations , ; therefore rules in a neural network . The are not to direct subject they manipulation operate at the level of individual processingunits. They specify how units compute their activation, how activation is transmitted to other units, and how connection weights are modified. Thus, representationsdo not act as a . Rather unit to affect other representations , individual processingunits interact with and causally affect other units. Intelligent behavior emergesfrom the way in which interacting units are connected ; it does not arise from the . direct interaction and transformation of representations This distinction between neural network and symbolic models has given rise to controversy over the need for explicit symbolic rules to account for certain cognitive activities. In particular, the controversy has centered around the need for explicit rules in models of natural languageprocessing : Prince : Pylyshyn, 1988; Pinker & , 1988). The study of language (Fodor & processinghas been viewed as a crucial test of the adequacyof neural network models . Traditionally, languagehas been viewed as the prototypical example of a . The system consistsof a finite number rule-governed, combinatorial system

Introduction to Neural Networks

of discrete elements (e.g ., phonemes , words, phrases ). These elements are combined systematically to create larger hierarchical structures (e. ., g phonemes combine to form words, words combine to form phrases , etc.). The rules that operate to createtheselarger structuresare sensitive to the syntax of their constituent elements (Fodor & Pylyshyn, 1988). This can be seen in a complex sentencesuch as "The classesthat the professor teachesare canceled ." This sentence has a constituent structure such that words combine to form phrases(e.g ., noun and verb phrases ). Thesephrasescombine further to form a sentenceembeddedwithin a .sentence . That is, the sentence" the " professor teachesthe classes is embeddedin the sentence" the classesare canceled ." The rules that operate to construct this embeddingmust be sensitive to the structural relations between the two sentences . For example , the " verb " teaches must agree in number with the noun in the embedding (i.e., ), not the noun in the main clause(i.e., classes professor ). The challenge in neural networks has been to implement complex structural relationships such as constituency in a system that lacks explicit rules specifying how structuresare to be combined (Lachter & Bever , 1988; Fodor & Pylyshyn, 1988; Pinker & Prince , 1988; Prince & Smolensky , 1997). Although progress has been made in modeling aspectsof language performance such as morphology, word recognition, and lexical semantics , it is far from clear that these models can learn to represent complex syntactic relations. , Elman(1993) has attempted to meet this challengeby training a Recently neural network to representsentences . He found containing relative clauses that a simple recurrent network learned to represent the hierarchical structure of such sentencesunder certain conditions. In particular, successwas achieved only when the network learned to represent simple sentences ' before it learned to represent complex ones . Elman s results are extremely promising. They suggest that neural networks have the representational power necessaryto capture complex structural relationships (Elman , 1990, 1993). However, the neural network approachdoes not yet provide the systematic overview of languageprocessing that would make it a satisfactory alternative to generativelinguistics.
The Nature of Neural Network Simulations

As we discussed above , one of the attractive properties of neural networks is their ability to discover a set of connection strengths that capturesthe internal structure of a domain. As an example , consider a reading model, called NETtalk, developedby Sejnowskiand Rosenberg(1987, 1988). NETtalk was trained to produce phonemesin responseto input in the form of English text. The network was constructed in three layers. It contained a level of input units, a level of output units, and an intervening level of "hidden units." The hidden units were densely connectedto both the input and output units. The network receiveda string of sevenletters and spaces of text as

t Levine: Neural Network Modeling ,& Long . Parks

input and was trained to produce the phonemecorrespondingto the letter in the center of the string. With sufficient training, the model discovered the . This implicit rules that govern the transformation of visual letters to sounds and words the set of to model the enabled training produce generalizebeyond the appropriate phonemic output for a large corpus of novel English words. Exactly what did NETtalk learn that enabled it to perform the reading ? Sejnowski and Rosenberg(1987) examined the pattern of activity distributed task ' across NETtalk s hidden units. They found that some units were . However, the activated by vowels; others were activated by consonants model had so many connectionsamong units (more than 18,000) that it was impossible to be more specific about how the network actually performed the task. to The NETtalk model illustrates a problem that has led some researchers question whether neural network models should be viewed as simulations of , 1988; , 1988; Fodor & Pylyshyn, 1988; Massaro cognitive theories (Estes Haan . For de 1994 van & 1991 , , ) ; Hezewijk example McCloskey McCloskey, (1991) has argued that neural network modeling is very different from traditional simulation. In a traditional computer simulation, the modeler specifies . The modeler knows that how eachfeature of a theory is to be implemented ' the modeler has the simulation instantiatesthe theory s assumptions , because . specifiedtheseassumptionsexplicitly The situation is very different in a neural network simulation. The modeler does not specify how the system is to implement the featuresof the theory. Rather , the system discoversits own implementation by meansof the learning " ), a neural network modeler algorithm. According to McCloskey (1991 " ' ' grows the network rather than building it (p. 391). This causesatremendous the performanceof networks as complex as NETproblem in assessing talk. Neural networks of any significant size are hugely underconstrained , , 1988; Reeke& Sporns , 1994; Fodor & Pylysh~ 1988; Massaro (Bullinaria 1993). The model discoversa solution that producesthe desiredoutput from the specifiedinput. However, it is likely that this solution is one of an infinitely ' large number of solutions to the problem. Does NETtalk s particular solution have theoretical implications that are different from alternative ? Is the knowledge encoded in the weights specific to the details solutions of this particular implementation (e.g ., seven characterinput strings, output constrainedto the middle letter)? Which featuresof the model implement the ? At present , theory and which are idiosyncratic to this particular simulation such answer to our knowledge of neural network functioning is insufficient . questions Given the difficulty in understanding how a complex network accom plishes a particular task, McCloskey (1991) has argued that neural networks should not be viewed as cognitive theories or simulations of cognitive . theories ; rather, they should be viewed as tools for theory development In particular , he has argued that networks can be thought of as analogs to

Introduction to Neural Networks

animal models of cognitive activities and disorders . Animal models are not simulations of cognitive theories ; rather, the animal system is an object of study. The goal is to develop a theory of its functioning that can then be extended to the human system . The theory may need to be modified significantly when it is applied to human functioning. However, the two systems may shareenough crucial featuressuch that a systematicanalysisof the animal . Similarly system aids in the development of a theory of the human system , neural networks can be studied to develop a theory of their structure and functioning. This theory can then be examined to determine the extent to which it can be applied to the human system . Viewed from this perspective animal , systemsand neural networks share some advantagesand disadvantagesas models of human functioning. An advantageof these systemsis that they can be manipulatedin ways that the human systemcannot. For example , thesemodels can be subjectedto tightly controlled training regimens or they can be lesioned to examine the effects of different types of damage . A disadvantageof these systemsis that they . may contain critical featuresthat have no counterpart in the human system The challenge will be to understand neural networks in sufficient detail to determine whether the process es that determine their performanceare similar to those that underlie human performance .
Modular vs . Interactive Architectures

and psychologistsbelieve that the architecturalsystem Many neuroscientists underlying human cognitive performance is functionally modular. That is, the system is composed of a number of dedicated processing components (i.e., modules ). Each component has a proprietary databaseand is informa(Fodor, 1983, 1985). This meansthat each component tionally encapsulated a ( ) contains knowledge that is specificto the operations that it performs and (b) this knowledge is available only to the component responsiblefor these . In addition, the partial results of operations that are specific to operations one component are unavailableto other components . to this view of the functional architecture , interaction among According modules is severely limited (Fodor, 1983, 1985). In particular, a module interactswith other modulesin the systemonly to receivethe computational end product from other modules or to communicateits own end product to thesemodules . For example , a module that is responsiblefor lexical analysis (i.e., identifying words from phonemic input) will receive input from amodule dedicatedto phonetic analysis . Once the lexical module has received this it will its own . As the lexical module performs its , input perform operations function, it will be unaffectedby information available from other modules , 1979; Garrett, 1981). This includes information that (Fodor, 1983; Forster , such as information about the semanticcontext in may be potentially useful which a word appears . Once the module has completed its function, it will present its product to the modules to which it is connected(e.g ., a syntactic

, & Levine: Neural Network Modeling Long , Parks

module). Thus, modules receive input from a small set of other modules and . produce a completedproduct in discretestages The assumption of modularity has implications for how neuropsychologists , 1994, and interpret the effects of brain damage (see Farah Nachson & Moscovitch, 1995, for a review). Brain damage can have very selective effects on cognitive abilities. Neuropsychologists often use information about the selective effects of damage to make inferencesabout the , 1981; Shallice , underlying functional architecture (Allport , 1985; Capian 1988). Theseinferencesare basedon an assumptionof " locality ." Damageto one component of the architectureshould have relatively local effects to the extent that the componentsin the system are modular, that is, information. Processingin undamagedcomponents should be affected ally encapsulated only to the extent that they receivedegradedinput from a damagedmodule. ' Because undamagedcomponentscontinue to function normally, the patient s impairment should relate in a fairly straightforward way to the behavior of the damagedcomponent . Researchers who work with neural network models have begun to question whether a dissociation of behavior after damage to the architecture , 1994; Farah& McClelland, 1991; (Farah implies a dissociationof mechanism Hinton & Shallice 1991 & Farah , ; Tippett , 1994). This question has arisen because many neural network models do not contain informationally encapsulated . The central properties of neural networks contrast components with the . First, the knowledge contained directly properties of modular systems in a neural network is distributed acrossa set of connectionweights; it is not localized to a particular component . Second , neural networks are interactive and processinformation in a graded and continuous fashion . Representations can be partially active either because some subsetof the units that comprisea representationare fully active or all of the units in a representation have some subthresholdlevel of activation. This partially active information can influencethe activation of units elsewherein the network. Thus, interaction is not limited to the endpoints of processing . Rather , the partial of one of the network are available to other products part parts continuously . during processing As the earlier section on adaptive resonance models suggested , many networks of that type exhibit a greater degree of modularity, with components that have distinct informational significance . For example , Levine and Prueitt 1989 modeled effects of cortex ( ) prefrontal damageon Wisconsin Card Sorting Test performance . Their model included componentsthat were labeledas " features " " " , and number of card designs (color, shape ), categories (of cards ), " reinforcement" " " " " , attentional blases , and habits. The model did not specify in detail exactly which parts of the brain correspondedto which components of the network. However, the network' s theoretical structure was rich enough to provide some potential hypotheses about what types of brain connections , and in which regions, might be present and account for the simulatedbehaviors (Hestenes , 1998; Levine , 1996).

Introduction to Neural Networks

Do neural network models exhibit functional dissociations of behavior ? Recent work suggests after damageeven when they are not modular systems Hinton and Shallice that they do. For example 1991 , ( ) trained a network to produce the appropriate semantic output for a word when it receivedletter strings as input. The network learned to discriminate visually similar input (e.g ., cat vs. cap , horsevs. hoarse ) as well as semanticallysimilar hat . e. . vs. vs. The network was fully interactive in , cap ) input ( g , peach apricot the sensethat it did not contain modules dedicated to a particular process , semanticprocessing ). Hinton and Shallicefound that (e.g., visual processing a lesion anywhere in the network produced an impairment similar to that seen in deep dyslexia. The model produced both visually based (e.g ., cap read as cat) and semanticallybasederrors (e.g ., cap read as hat). In addition, the model produced more combined visual and semanticerrors (e.g ., cat read as rat) than would be expectedby the baserates of eachtype of error alone. Farahand her colleagueshave argued that damageto one part of a system can have profound effectson the functioning of undamagedparts of the system , 1994, for a review). This claim is supported by simulationsof (seeFarah the effects of damage on network models of semantic memory (Farah & -Schreiber McClelland, 1991), visual attention (Cohen, Romero , et aI., , Servan ' & Vecera O 1993 1994), and visual face recognition (Farah , ). For Reilly, , Tippett and Farah (1994) trained a network to produce semantic example ' in response to visual input. They then damaged the network s output semantic representationsto simulate the type of damage suffered in Alz ' . They found that the performanceof the model as a whole heimers disease was affected when factors such as word &equency or visual degradation . increasedthe processingdifficulty of any of its components The simulations discussedabove suggest that (a) damageto one part of a system can affect the functioning of undamagedparts and (b) behaviors that are functionally modular (e.g., visual vs. semantic processing ) can be performed by a system whose structure is not modular in the sameway. These findings are controversial partly becausethey have such profound implications for cognitive neuropsychology. They suggest that it may bemisleading to make inferencesabout the function of a damagedcomponent on the ' . Dissociations of behavior may tell us little basis of a patient s symptoms . about the underlying functional architecture CURRENT LIMITATIONS OF NEURAL NETWORK MODELS Current network technology limits the application of neural networks to . Restricting the size of a network to the minimum small versions of real problems number of units and connectionsneededto solve the problem allows , and period of time, improves generalization training to occur in a reasonable ' makes it easier to understand the model s representationaland processing properties. However, small systemsare likely to differ &om large systemsin important ways. In order to scaleup networks to sizescomparableto those

: Neural NetworkModeling , Parks , & Levine Long

found in the brain, severaltheoretical and computational limitations must be overcome (Bullinaria : Chater , 1994 ; Bullinaria & : , 1993; Feldman , Fanty, & Goddard, 1988).

TheProblem of Internal

Structure

The first challengein solving the scaling problem involves finding ways to constrain the searchspaceduring learning. Systemsthat lack built -in knowledge about the world learn incredibly slowly and are overly dependent on , a knowledgeable tutor , or carefully particular input-output representations trials Bullinaria : Ackley, , 1994; Hinton , Sejnowski ( , & sequencedtraining 1984). Shepard(1989), in particular, has claimed that nontrivial learning can occur only in a system with an internal structure that captures the regularities in its environment (Shepard , 1987a ). Without this structure , a system will have no basisfor generalizing information from one context to another. , this internal structure is acquiredas organismsevolve in an environment Presumably that is invariant across certain dimensionsS ".epard (1989) suggests severalgeneral constraints that may be acquiredby terrestrial species ; these include knowledge that " material objects are conserved , that spaceis locally Euclidean and three-dimensional with a gravitationally conferred unique upright direction, and that periods of relative light and warmth alternate with periods of relative dark and cold in a regular cycle" (p. 107). Organisms that come equipped with knowledge about the universal constraints of the world have an adaptive advantage (Shepard , 1987b). These constraints narrow the set of behaviors that must be learned by trial and error. Shepard(1989) provides severalexamplesof the functions that may be served by such internalized constraints . One of these functions may be the to an . Consider a ability recognize object regardlessof its position in space system with sensory or input units that become activated in response to . The system also has a set of "hidden light reflected off an object in space " units that will come to represent the object after it has been presented repeatedly in more or less the sameposition. Now imagine that the object is moved in spacesuch that it no longer functions to activate the sameset of , the system may no longer recognize sensory units. In current network models this as the object that it had learned previously. That is, information learnedabout the object in one location in spacemay not transfer when the object is moved to a new location. One solution to this problem is to train the system by presenting the object in all possible spatial locations and orientations. However, this would require a prohibitive amount of training. This solution is even less desirable when you consider that such training would be required for each class of . object that the systemmight be required to recognize 1989 two of how neural ( networks might ) Shepard suggests possibilities be designedwith the internal structure necessary for effective generalization from learning. First, the internal constraints may be hardwired into the sys-

Introduction to Neural Networks

. Second tern prior to any learning experience , neural networks may be . evolved over time to capture the invariant properties of their environments in either of these obstacles However, we face enormous suggestions implementing given our current knowledge of the microarchitectureof the human . systemand the evolutionary principles that have guided its development The Problem of Biological Plausibility The second challenge in neural network modeling is to develop network architecturesand learning mechanismsthat better represent those found in : Carey, 1992). Modelers real biological systems (Crick, 1989; Eagleson & generally agreethat input output relations found at one level of organization can be simulatedby an inAnite variety of models at lower levels of organization : Sporns : Pylyshyn, 1988; Pylyshyn, 1984; Reeke & , 1993; (Fodor & neural networks it is that this reason . For 1990 , , ) represent important Shepherd . properties found in real biological systems Currently, neural network models often incorporate features that are . For example known to be biologically questionable , many models have . in the same set of connections es realized and inhibitory process excitatory output either or be of That is, a unit may activating inhibitory capable sending . This is not a property found in real neural to other units in the system anatomy. A neuron may receive both excitatory and inhibitory input, but it sendsonly one type of signal. In addition, many networks incorporate back criticized that has been a mechanism mechanism as a , learning propagation : Carey, , 1989; Eagleson& repeatedly for its biological implausibility (Crick 1992). The backpropagation learning mechanism requires that the connections between units be able to convey signals in two directions. This is to the equivalent of an axon with the capability of both sending messages has shown in return. Grossberg(1987) other neuronsand receiving messages that the neural hardwareneededto implement such a mechanismcontradicts es. our knowledge of brain organization and process es that better reflect our and structures network process Developing understanding of real biological systems should help neural network modelers develop a priori answersto questionssuch as: (1) How many units and ? (2) how many layers should a network contain to perform a particular task ? (3) How many training trials Which learning rule should be implemented of the domain should be ? (4) What characteristics should the network receive ?, and (5) If the model is to be captured by the input output representations and how many? be units should which , damaged damaged
The Problem of Catastrophic Interference

A third limitation of current network models is that they tend to exhibit a " " : Cohen, 1989; interference (McCloskey 81 phenomenoncalled catastrophic ' Ratcliff 1990 1995 on 81 : O Mc McClelland, ; , ). Catastrophic Reilly, Naught ,

: Levine: Neural Network Modeling ,& Long, Parks

interferenceoccurs when a model exhibits rapid forgetting for well-learned items when it is trained on a new set of items. This occurs primarily in situations involving the rapid acquisition of arbitrary associationsbetween are required to model a wide variety of inputs and outputs. Suchassociations memory phenomena(e.g ., memory for items in a list, memory for names , memory for telephonenumbers ). McCloskey and Cohen (1989) demonstratedcatastrophicinterferencein a neural network model of a paired-associatelearning task. In the pairedassociate task , participants learn associations between arbitrarily paired words or nonsensesyllables . They begin by learning one list of associations -responsepair of words called the AB set. The AB set is a stimulus e. . ( g , book pear , drawer car). Participants are trained on the AB items until they accurately produce B (e.g ., pear ) when they are given A (e.g ., book). A typical interference manipulation is to present participants with a second set of words (AC ) once their performanceon the AB set has reached asymptote. The AC set involves the same list of stimulus words now paired with a different list of response words (e.g ., book-mouse , drawerchurch ). Participants exhibit a gradual loss of the ability to retrieve the AB list after training on the AC list. However, AC learning does not completely interfere with AB retrieval. Participants exhibit about 50 percent correct performance on the AB items even after AC learning has reachedasymptote . and Cohen (1989) modeled the paired-associatelearning task McCloskey in a backpropagationnetwork. The network consistedof three layers: a layer of input units, a layer of output units, and an intervening layer of hidden units. A subsetof the input units representedA items; another subsetrepresented the B and C items. The network was trained first on the AB items and then on the AC items. They found that the model's performance differed . In particular, the dramatically from the performanceof human participants model lost all ability to retrieve the AB items after only a few trials on the AC list. Ratcliff (1990) observed similar catastrophicinterferencein a neural network model of recognition memory. The model tended to track the information presented last in a list, exhibiting rapid forgetting of welllearneditems that had been presentedearlier. Researchers have found that catastrophicinterference can be avoided by the connection adjusting weights to reduce the overlap between competing French , 1991, 1992; Hetherington & ( : Seidenberg responses : , 1989; Sloman& Rumelhart , 1992). However, French(1991) has argued that reducing overlap to avoid catastrophic interference comes at the expenseof generalization . Generalization occurs in neural network models precisely becausepatterns that representsimilar information overlap. The challengewill be to develop meansof reducing interferencethat do not also reduce the exploitation of shared structure . Carpenter (1997) has developed a mathematical theory that the appropriate choice of learning laws for connection suggesting

Introduction to Neural Networks

CONCLUSIONS
The neural network field has matured rapidly since it was introduced to mainstream neuroscientists and psychologists in the early 1980s. The field has moved from exploring the properties of different classes of neural network models to testing these models as alternatives to existing theoretical frameworks . We have briefly described some of the basic principles of neural network modeling and the properties that make them an attractive alternative to other approach es. We have also reviewed some of the theoretical and computational limitations of current models . As we discussed, neural network modelers face enormous problems in developing models that are more biologically plausible , have the preexisting structure necessary to engage in nontrivial leaming , and provide adequate accounts of higher - level cognitive processes such as language and reasoning . Although these are formidable obstacles, they provide an exciting set of theoretical and computational challenges to current and future researchers. In spite of methodological and theoretical differences between backpropagation and adaptive resonance programming , preliminary modifications in describing programming parameters have produced some overlap in how processing elements and measures may be interpreted . For example , Cardoso and Parks in chapter 9 use backpropagation , whereas Parks and Levine (in chapter 15) utilize adaptive resonance programming . Yet the outcomes described in both chapters have produced significant advances in neurocomputational and neuropsychological understanding of executivefunctioning (Rabbitt , 1998 ). We hope to encourage clinicians to begin their own exploration of com putational neuropsychology . An essential step, particularly for those clini clans who are unfamiliar with the methodology , is to acquire backpropagation or ART software , or both . Clinicians may choose to be part of a team of investigators . Such teams typically consist of a mathematician or computer scientist fluent in computer programming techniques , a neuropsychologist , and a physician (psychiatrist or neurologist ). Access to clinical populations ' (e.g ., Alzheimer s, head trauma , discrete lesions , and others ) and control subjects , both of whom have undergone quantitative neuropsychological batteries , is needed to integrate a priori theories with empirical psychometric data. Neural network modeling of complex neuropsychological data is still at an early stage of development . Current research involving clinical populations ' attempts to approximate , or in some cases replicate , patients performance relative to the performance of a control group . Institutions that have access to large databases of neuropathological conditions with corresponding neuropsychological data can, with the appropriate personnel and

: Neural NetworkModeling : Levine ,& Long , Parks

, begin utilizing computational methodology. Institutions computing infrastructure with access to neuroimaging data (computed tomography, magnetic resonance imaging, functional MRI , single photon emission computed ) are at a particular advantage , since tomography, electroencephalography thesedata may both guide in the development of computer models and provide validation when the computer program results are consistentwith neurophysiological and neurostructural information .

NOTFS
1. Substantialportions of this section were adaptedfrom Parksand Cardoso (1997). Used with

. pennission

REFEREN CFS
). Distributed , modular . In S. K . Newman , anddysphasia Allport, D. A. (1985 memory subsystems - 452 & : R . Epstein " entperspectives .), Cu in dysphasia (Eds ill (pp. 421 ). New York: Church . Uvingstone Amari of randomly , S. I. (1971 connected threshold element ). Characteristics networks andnetwork . Proceedings , 59 , 35- 47. systems of theIEEE Anderson modelutilizingspatial , J. A. (1968 correlation ). A memory functions . Kyberne storage - 119 tik, 5, 113 . Anderson for memoryorganization , J. A. (1970 ). Two models traces . Mathematical usinginteracting - 160 Bioscienas . , 8, 137 Anderson : Hinton , J. A , & , G. E. (1989 in the brain . In G. E ). Modelsof information . processing Hinton & : J. A Anderson .), Parallel models (Eds : (pp. 23- 62). Hillsdale , NJ of associative memory Erlbaum . Anderson : Kushmerick , J. R., & . N. (1992 . In J. RAnderson ). Towerof Hanoiandgoalstructures - 142 : Erlbaum . (Ed ), Rules , NJ ). Hillsdale of themind(pp. 121 Anderson : Rosenfeld . (Eds .) (1988 , J. A., & ,R . Carn ). Neurocomputing : Foundations of research : MIT Press . , MA bridge Buninaria . A. (1994 nonlocal : Rules ,J of thegame . Behavior andBrainSciences ). Simulating systems , 17 , 61- 62. Bu Ulnaria : Chater in artiAdalneuralnetworks , J. A., & , N. (1993 : Implications ). Doubledissociation for neuropsychology . In Proceedings annual Hve Science of thefifteenth conference of theCogni - 264 : Erlba \Ul '\. (pp. 227 , NJ ). Hinsdale Society Burton : Bruce . AM .. & . V. (1994 ). Localreoresentations without thelocalityassumption . Behav
loral and Brain Sciencl S , 17 , 62- 63.

, D. (1981 localization of linguisticfunctions : Logicaland empirical ). On the cerebral Capian issues de Adt analysis andfunctional localization . Bra -in tmdLanguage - 137 . , 14 , 120 surroonding Cardoso . . 1991 the Towers of Hanoi . AI ,J ( ) Revisiting , October , 49 53. Erperl , G. A. (1997 ). Spatial Carpenter patternlearning , catastrophic forgetting , and optimalrulesof transmission . In D. S. Levine & : W. R. Elsberry .), OpHma -lity in biologiC R I andRr Rl (Eds synaptic Hfici networks ?(pp. 288 - 316 : Erlbawn . ). Mahwah NJ

Introduction to NeuralNetworks

: Grossberg . (1987 architecture for a self , G. A , & - organizing ). A massively Carpenter ,S parallel neural mamme . Computer Vision . , Graphics , andImage , 37 , 54- 115 recognition pattern Processing

Caudin : Butler . Cambridge , M., & , C. (1990 . ). NRtur , MA: MIT Press Rily intelligent systems Chauvin : Rumelhart . (1995 : Theory , Y., & , D. E . ). Backpropagation , architectures , andapplications : Erlbaurn Hinsdale . , NJ
-Schreiber Cohen : Farah , J. D., Romero of spatial , R. D., Servan , D., & , M. J. (1994 ). Medtanisms attention : Therelation of macrostructure to microstructure in parietal . ] oUn I431 of Cognitive neglect - 387 Neuroscience . , 6, 377 of a mentalanalog of an external rotation . Perception , L. A (1976 and ). Demonstration Cooper --302 . , 19 , 296 Psychophysics - 132 exdtement aboutneural networks . Nature . Crick , F. (1989 ) Therecent , 337 , 129 ., & : Carey networksdo not model brain function . ,R , D. P. (1992 ). Connectionist Eagleson - 735 Behavioral andBrain Sciences . , 15 , 734 Elman structure in time . Cognitive Science . - 212 , J. L. (1990 ). Finding , 14 , 179 Elman and development in neuralnetworks : The importance of starting , J. L. (1993 ). Learning - 99. . Cognition small , 48 , 71 -processing Estes for combining connectionist and symbol , W. K. (1988 ). Towarda framework models . ]O Unl431 and 27 - 212 . , , 196 of Memory Language
Farah . M . J. (1994). Neuropsymological inference with an interactive brain: A cinque of the ' " and Brain Sciences , 17 , 43- 61. localityassumption Behavioral

: McClelland modelof semantic : Farah , M. J., & , J. L (1991 ). A computational memory impairment -sped -sped . oumlll mental : General Modality ] of Eldty andemergent category Erpm ' Psychology Eldty - 357 . , 120 , 339 ., O'Reilly : Vecera . P. (1993 overt andcovertrecognition as Farah , M. J , R. C., & ,S ). Dissodated - 588 anemergent neural networks . Psychological Review . , 100 , 571 propertyof lesioned : Ballard modelsand their properties . Cognitive Feldman , J. A., & , D. H. (1982 ). Connectionist - 254 Science . , 6, 205 : Dell, G. S . (1985 connectionist Feldman , J. A., Ballard : , D. H., Brown . C. M., & ). Rochester papers 1979 - 1985 . Technical . Rochester of Computer Science of , NY: Department , University reporil72 Rodtester . : Goddard with structured neural . Feldman , J. A., Fanty networks , M. A., & , N. (1988 ). Computing IEEE . , 21 , 91- 103 Computer Fodor . Sussex . : Harvester Press . , J. A. (1976 ). The , U.K language of thought Fodor . Cambridge . , J. A. (1983 , MA.: MIT Press ). The modularity of mind Fodor of themodularity of mind . Behavioral andBrain Sciences , J. A. (1985 ). Preas , 8, 1- 42. Fodor : Pylyshyn andcognitive architecture : A aitical analysis , J. A., & . ). Connectionism , Z. (1988 . , 3- 71 , 28 Cognition Forster and the structure of the language . In W. , K. (1979 ). Levelsof processing processor : E. Walker(Eds .), Sentence studies to Merrill Garrett : Psycholinguistic Cooper& processing presented : Erlbaum . , NJ (pp. 27- 85 ). Hillsdale . M. (1991 -distributed to overcome French , R ). Usingsemi representations catastrophic forgetting in connectionist . In Proceedings networks annual science of thethirteenth cognitive conference - 178 : Erlbaum . (pp. 173 , NJ ). Hillsdale

: Levine : Neural NetworkModeling ,& Long , Parks

French . nectionist Garrett

R . M . ( 1992 networks , M . ( 1981 ,

) . Semi - distributed

. Connt ) . Objects

representations Ch' on Science , 4 , 365 - 377 . of psycholinguistic

and

catastrophic

forgetting

in

con -

inquiry of

. Cognition learning

- 101 . , 10 , 97 with .

Grossberg Journal Grossberg development Grossberg Grossberg Cognitive " Grossberg Hanson

S . ( 1969

) . Embedding C43l Psychology ) . Adaptive of neural

of M Rthtmati , S . ( 1976 and

fields : A theory - 239 . , 6 , 209 classification detectors

physiological

implications

pattern feature and

and . Biologi

universal C43l Cybtrnttics

recoding

: Part

I . Parallel

coding

- 134 . , 23 , 121 : Reidel to Press . resonance .

, S . ( 1982

) . Studies

in mind

bra -in . Dordrecht : From

, Netherlands activation

, S . ( 1987 ) . Competitive Science , 11 , 23 - 63 . , S . ( 1988 ) . Neura - l networks

learning

interactive

adaptive

and

na- tuml

intelligence

. Cambridge

, MA

: MIT and

Press .

, j . S ., &: Burr , D . j . ( 1990 networks

) . What

connectionist

in connectionist Hebb

. Bth Rviora -1 and Ziltion

models learn : Learning Bra -in Sciences , 13 , 471 - 518 . Rvior . New in York mental : Wiley . . In

representation

, D . O . ( 1949

) . The orgtmi

ofbth

Hestenes

, D . O . ( 1998

) . Modulatory

mechanisms

disorders

D . j . Stein

&: j . Ludik University

( Eds . ) , Neura -l networks Press .

in psy Chopa -thology

- 164 . ( pp . 132 ) Cambridge

, U .y _ : Cambridge

" " in con : Seidenber , P. A " & interference ~ M. S. (1989 ). Is there catastrophic Hetherington nedionist networks ? In Procudin Science ,rs of the Cognitive . 2633 . : (pp ) Hinsdale , NJ Society Erlbaum . Hinton ., McClelland . L, & : Rurnelhart , G. E ,J . In DE . , D. E. (1986 ). Distributed representations Rurnelhart & : J. L McClelland .), Par Rllel distributed : Erplorations in themicrostructure (Eds pr0C5ing - 109 . , Vol. 1 (pp. 77 ). Cambridge , MA: MIT Press of cognition Hinton : Sejnowski , G. E., & andrelearning in Boltzmann , T. J. (1986 . In DE . machines ). Learning Rurnelhart & : J. L McClelland .), Parallel distributed in themicrostructure (Eds pr0C5ing : Erplorations - 317 , Vol. 1 (pp. 282 . , MA: MIT Press of cognition ). Cambridge Hinton : Ackley , G. E., Sejnowski , T. J., & Machines : Constraints , D. H. (1984 satisfaction ). Boltzmann networksthat learn . Technical CMU-CS -84 - 119 : Department of , Pittsburgh Report -MellonUniversity Science . , Carnegie Computer Hinton : Shall ice , G. E., & an attrador network : Investigations , T., (1991 of acquired ). Lesioning - 95. .ia, Psychological Review , 98 , 74 dyslex andphysical , J. J. (1982 with emergent ). Neuralnetworks collective Hopfield systems computa - 2558 tionalabilities . Proceedings . , 79 , 2554 of theNational Academy of Sciences Koch : Segev in neuronal , C., & , I. (1998 ). Methods .). (2nd ed modeling : Fromionsto networks . , MA: MIT Press Cambridge -theoretical : A system Kohonen , T. (1977 . New York : Springer ). Associative memory approach . Verlag -Verlag Kohonen -organi Z /ltionandassociative . Springer , T. (1984 . ). Self memory Lachter : Bever , J., & , T. G. (1988 andtheories of language ). The relationbetween linguisticstructure : A constructive of some connectionist . Cognition , 28 learning , critique learningmodels - 247 195 .
Levine , D. S. (1983 ). Neural population modelin RtiCill Bio~ and psychology: A review. Mathem sciences , 66, 1- 86.

Introduction to Neural Networks

. (1989 Levine networks . AI Erptri , D. S , December ). Thethird wavein neural , 26- 33. Levine andcognitive . Hillsdale to neural : Erlbaum . , D. S. (1991 ). Introduction , NJ modeling Levine of the prefrontal . In J. Reggia executive , D. S. (1996 ). Modelingdysfunction , E. system - 439 : R. Berndt(Eds .), Neuralmodeling : ,& , NJ (pp. 413 ). River Edge of braindisorders Ruppin World Scientific . Levine : Prueitt effects of frontallobedamage : Novelty and , D. S., & , P. S. (1989 ). Modelingsome - 116 . Neural . Networks , 2, 103 perseveration Massaro aiticismsof connectionist models of human . Journal , D. (1988 ). Some of performance - 234 andLanguage . , 27 , 213 Memory McClelland : O'Reilly , J. L., McNaught , R. C. (1995 on , B. L , & ). Why therearecomplementary in the and : Insights from the success esandfailures of systems learning hippo campus neocortex - 457 connectionist models of learning andmemory . Psychological Review . , 102 , 419 McClelland . L , Rumelhart : the POPReseareh distributed ,J , D. E., & ). Parallel processing Group(1986 in the microstructure andbiological . models , Vol. 2: Psychological : Explorations of cognition . , MA: MIT Press Cambridge andtheories : Theplace . of connectionism in cognitivescience , M. (1991 ). Networks McCloskey - 395 Science . , 2, 387 Psychological : Cohen in connectionist networks : The interference , M., & , N. J. (1989 ). Catastrophic McCloskey . In G. H. Bower Ed . The and motivation Vol. , ( ), of learning learning psychology sequential problem - 165 24 (pp. 109 : Academic Press . ). NewYork ., & : Pitts imminent in nervous activity of the ideas McCullochW.5 , W. (1943 ). A logicalcalculus - 133 . Bulletin Rtical . , 115 of Mathem BiophysicsiS Mervis : Rosch of natural . Annual Review , C. B., & , E. (1981 , ). Categorization objects of Psychology 32 . , 89- 115 Nachson : Moscovitch .). (1995 issue , I., & , M. (Eds ). Modularityandthe brain[special ]. Journal of andExperimental . Clinical , 17 Neuropsychology - 183 Newell . Cognitive Science . , A. (1980 , 4, 135 ). Physical symbolsystems Newell level . Arlijidal Intelligence . , A. (1982 , 18 , 87 127 ). Theknowledge Newell ., & . (1989 : Laird architectures for cognition . In M. I. , A. Rosenbloom , P. S , J. E ). Symbolic Posner .), Foundations science . Cambridge . , MA: MIT Press (Ed of cognitive -Wesley Pao andneural . networks . New York : Addison , Y. H. (1989 ). Adaptive recognition pattern Parks : Cardoso . (1997 : distributed and executive , R. W., & ,J ). Parallel functioning processing Towerof Hanoineural . International networkmodelin healthy controls andleft frontallobepatients - 240 . , 89 , 217 Journal of Neuroscience Parks . W., Long ,R , D. S., Crockett , D. J., McGeer , P. L., Dalton , I. , E. G., McGeer , D. L , Levine E., Zec . M. (1991 : Bower , R. F., Becker , R. E., Coburn , K. L , Siler , G., Nelson , M. E., & ,J ). Parallel distributedprocessing and neuralnetworks : Origins and cognitivefunctions . , methodology - 214 International . , 60 , 195 Journal of Neuroscience -time representation : Uinas in the brain : The cerebellum as a Pellionisz , A., & , R. (1982 ). Space time . metric tensor Neuroscience 7 2249 2970 . , , predictive space Pinker : Prince andconnectionism : A parallel distributed , S., & , A (1988 ). On language processing modelof language . Cognition . , 28 , 73193 acquisition Posner : Keele . W. (1970 of abstract ideas . Journal , M. I., & ,S ). Retention of Experimental Psychology - 308 . , 83 , 304

: Levine : Neural NetworkModeling ,& Long , Parks

Prince : Smolensky : HornNeuralnetworks , A., & , P. (1997 to universal . ). Optionality grammar - 1610 Scitn Ct . , 275 , 1604 andCognition . Cambridge . ). Computation , MA: MIT Press Pylyshyn , Z. (1984 in cognitivescience . In M. I. Posner , Z. W. (1989 .), Foundations ). Computing Pylyshyn (Ed of science . 51 91). Cambridge . ( , MA: MIT Press cognitive pp . In M. Minsky (Ed), Semantic , M. R. (1968 ). Semantic Quillian memory information processing - 270 . (pp. 227 ). Cambridge , MA: MIT Press Rabbitt . Bristol , P. (1998 : Francis . ). Methodology , PA: Taylor& offrontalandexecutive function : DoverPress , N. (1960 . ). Mathematical , Vol. 2. NewYork Rashevsky biophysics Ratcliff . (1990 models of recognition ,R : Constraints ). Connectionist memory by learning imposed - 308 andforgettingfunctions . Psychological Review . , 97 , 285 Reeke ., & : Spoms , G. N., Jr basedmodelingand computational , O. (1993 ). Behaviorally esto neuroscience . Annual Review - 623. , 16 , 507 of Neuroscience approach Rolls : Treves networks andbrain . New York , A. (1998 : Oxford University , E., & ). Neural function Press . Rosenblatt : Ptrceptrons andthetheory , F. (1962 . ). Principles of neurodynamics of brainmechanisms Books . , DC: Spartan Washington Rumelhart : Williams internalrepresentations , D. E., Hinton , G. E., & , R. J. (1986 ). Learning by error propagation . In DE . Rumelhart . L McClelland & : J .), Parallel distributed (Eds processing : - 362 in themicrostructure . , Vol. 1 (pp. 318 , MA: MIT Press ). Cambridge Uplorations of cognition Rumelhart : McClelland distributed , D. E., & in the , J. L (1986 ). Parallel : Uplorations processing microstructure . Cambridge . , Vol. 1: Foundatons , MA: MIT Press of cognition : Abelson structures : Information in mindand Schank , R. C., & , R. P. (1977 ). Conceptual processing machine . Mahwah : Erlbaum . , NJ aboutcomputation in cerebral cortex . In J. A. McClel , T. J. (1986 ). Openquestions Sejnowski land : the POPResearch .), Parallel , D. E. Rumelhart distributed ,& : Uplora Group(Eds processing tionsin themicrostructure Vol. 2: and models . 372 389 , (pp ). of cognition Psychological biological . , MA: MIT Press Cambridge : Rosenberg networks that learnto pronounce , T. J., & ). Parallel Sejnowski , C. R. (1987 English - 168 text. Complex . , 145 ,1 Systems : Rosenberg . (1988 andrepresentation in connectionist , T. J., & models . ). Learning Sejnowski , C. R In M. Gazzaniga .), Perspective in memory research andtraining . Cambridge . (Ed , MA: MIT Press Shallice in planning . Philosophical , T. (1982 Transactions ). Specific of theRoyal impairments Society London 298 199 209 . , , of Shallice to mental structure . Cambridge , T. (1988 University ). Fromneuropsychology , U.K.: Cambridge Press . law of generalization for psychological , R. N. (1987a . science ). Towarda universal Shepard - 1323 Science . , 237 , 1317 . N. (1987b of a mesh between ,R of the mind andregularities of ). Evolution Shepard principles the world. In Dupre .), Thelatest on thebest : Essays on evolution andoptimality , J. (Ed (p. 110 ). . , MA: MIT Press Cambridge of universal , R. N. (1989 : A challenge for con ) Internalrepresentation Shepard regularities nedionism . In L Nadel :R . M. Hamish(Eds .), Neural , LA . Cooper , P. Culicover connections ,& , - 134 mental . (pp. 104 , MA: MIT Press ). Cambridge computations

Introduction to Neural Networks

. N., & : Cooper Orm Rtions . Cambridge .R ). Mentlll imRgtSRndtheir trRnsf : , LA . (1982 , MA Shepard MIT Press . : Metzler -dimensional , R. N., & , J. (1971 ). Mentalrotationof three , 171 , objeds., Science Shepard - 703 701 . of realneuronarchitectures for neuralnetworksimulations , G. M. (1990 ). The significance Shepherd - 118 . In EL . Schwartz . ), Compuiati O MIneuroscience (Ed , MA: MIT (pp. 105 ). Cambridge Press . .A , & : Rumelhart interference in distributed memories Sloman , S , D. E. (1992 ). Reducing through . In Healy . Kosslyn :R . Shiffrin(Eds .), Essays in honor : K . EsttS ,S , Vol. 1, , & gating of w episodic - 248 : Erlbaum . , NJ (pp. 227 ). Hillsdale of connectionism . Behavioral andBrainSciences , P. (1988 ). On the propertreatment , Smolensky 11 , 1- 23. : Neuralcomputation . In L connections , P. (1989 ). Connectionist Smolensky / mental modeling Nadel :R . M. Hamish .), Neural connections , L A Cooper , P. Culicover ,& , mental (Eds computations . , MA: MIT Press (pp. 49- 67 ). Cambridge of extension . , P. (1990 , representation ). Problems , and computational irreducibility Suppes - 508 Behavioral andBrainSciences . , 13 , 507 : Farah modelof namingin Alzheimer ' s disease : , L. J., & , M. J. (1994 ). A computational Tippett ? Neuropsychology . , 8, 3- 13 Unitaryof multipleimpairments VanHezewijk ., & : de Haan . F. (1994 brainor the invisible hand ? Behavior ,R , EH ). Thesymbolic - 86. andBrainSciences , 85 , 17 Von derMalsburg - organization of orientation sensitive cellsin the striate cortex . ). Self , C. (1973 14 85 100 . , , Kybernetik Werbos : New toolsfor prediction andanalysis in thebehavioral , P. J. (1974 ). Beyond regression sciences . Unpublished PhiD . dissertation in , HarvardUniversity , Cambridge , MA Reprinted Werbos : From ordered derivatives to neural and networks , P. J. (1993 ). Theroots of backpropagation . NewYork : Wiley. forecasting political Widrow : Hoff, M. E. (1960 . In Institute , G., & , ). Adaptiveswitchingcircuits of Radio Engineers Western Electronic andConvention Show record , convention , part4. pp. 96- 104

: Levine: Neural Network Modeling ,& Long , Parks

Functional

Cognitive

Networks

in Primates

Ashford , Kerry L. Coburn , and J. Wesson M . Fuster Joaquin

The information-processing capability achieved by the human brain is a marvel whose basisis still poorly understood . Recentneural network models invoking parallel distributed processing have provided a framework for : the ,& appreciatinghow the brain performs its tasks(McClelland, Rumelhart POP ResearchGroup, 1986; Parks et al., 1989, 1992; Bressler 1995 . , ) The of distributed in nonhuman concepts parallel processingdeveloped primates provide useful models for understandingthe extraordinary processingcapability achievedby the human brain (Ashford, 1984; Goldman-Rakic , 1988). The field of neuropsychology can use this understanding to improve the , capability of assessing specifichuman cognitive functions suchas perception memory, and decisionmaking. The nervous systems of nonhuman primates provide clues to the brain systemswhich support human cognition. Monkeys can learn sophisticated , and in doing so they use structural and functional brain systems cognitive tasks . The functions of these systems highly similar to those usedby humans are revealed through depth electrode recording of single or multiple neuronal unit activity and event-related field potentials, and the anatomical distributions of the systems may be seen using high-resolution structural . However, the neural basesof cognitive scanningand histological techniques function becomemore clear when these techniquesare applied in a context in which a specific neuropsychological function is occurring. For example , when neurons in a monkey' s visual associationcortical region are observed to respond in the context of a visual memory task , the roles of both the neurons in that region and the region as a whole neural network appear to fall into a comprehensible framework. In turn, models of information processing in developed regions of the nonhuman primate brain have direct applicability to the function of analogousstructuresin the humanbrain (for reviews, seeFuster , 1995, 1997a ,b).

SYSTEM BUILDING BLOCKSOFTHE NERVOUS


Severalbasic principles of nervous system organization form the basis for , 1990). The adaptive understanding higher primate brain function (Jones

pathways have developed redundant and parallel channelsto assurethe reliability and fidelity of transmitted information, as well as to increase the and . Neurons and neural networks also have reliability of processing speed means for developed abstracting , retaining, and later retrieving information - the basic timespanning operations of memory. Progressively more , with higher levels of neurons complex levels of analysis form a hierarchy and networks performing progressively more complex information analyses and more refined responseproductions (Hayek, 1952). However, one general principle is: the more neurons involved in processing , the more complex the , 1991). But a larger number of potential analysisof the information (Jerison neurons also has a larger energy cost that must be borne by the organism and species , and hence a large brain must have a cost-benefit justification. Further there is a need for both functional specialization(e.g ., analysisof line , orientation or color) and generalization (e.g ., determining abstract relations between stimuli) of networks. NEURONS AND NEUROTRANSMITTER SYSTEMS The fundamental computational building block of the brain is the neuron, which contains dendrites for the input of information and an axon for the dissemination of the results of the neuron' s analysis . Typical invertebrate neural systems control muscle fibers by an excitatory acetylcholine neuron opposed by an inhibitoryy -aminobutyric acid (GABA ) neuron. In the vertebrates , acetylcholine neurons also work as activators throughout the nervous system , exciting muscle fibers and other effectors peripherally and numerous other systems centrally, including motor pacing systems activating in the basal ganglia and memory storage systems in the cortex. The GABA neurons of vertebrates presently are found only in the central nervous system where they still play the major inhibitory role &om the spinal cord up to the cortex. Serotonin neurons appear to mediate sensitization : Kandel , 1995), and serotonin neurons conditioning in the invertebrate (Bailey & , with the most widely distributed axons in the vertebrate brain, are retained in vertebratesfor a variety of central functions which require a conditioning : Azmitia, 1992). Similarly, catecholamine component (Jacobs & neurons developed in invertebrates , and playa role in reward-related learning in vertebrates(Gratton & : Wise, 1988). The principal neuron in the cerebral cortex is the pyramidal cell, which uses the amino acid glutamate as its neurotransmitter . Glutamate mechanisms are highly active in the olfactory system (Kaba , Hayashi , Higuchi, et al., 1994; Trombley & : Shepherd , 1993), and playa role in the analysesof chemicalstimulants . Olfactory functions include attending to and identifying a particular scentpattern, evaluating its significance , and retaining a memory

Introduction to Neural Networks

trace of the scent in its context. The major structural basis for information processingin the cortex may initially have developed in the olfactory system to serve this function. Hence , glutamate neuronsdeveloped their central role in the cortex, perceiving and retaining sensory information and making in the olfactory system . decisionsabout approachresponses The olfactory system may be thought of as a long-range component of the gustatory system , and the interaction of olfaction and gustatation can produce what is perhapsthe most powerful form of learning. The olfactory system itself can mediate aversive learning, but it is not particularly powerful . Aversive learning mediated by the gustatory system , however, can be A can be conditioned in a . taste avoidance extremely powerful response single trial and is unusually resistant to extinction. The interaction of olfactory and gustatory systemsis seenwhen odor and taste stimuli are combined ; -potentiated odor stimulus then acquires the same extraordinary the taste one-trial conditioning and resistance to extinction as the taste stimulus -Rattoni, Coburn, et al., , Kiefer, et al., 1984; Bermudez (Coburn, Garcia 1987). Although these phenomenawere discovered and have been studied -potentiated odor aversion learning in animals , both taste aversion and taste . They probably are seen in humans undergoing chemotherapy for cancer representa very specializedform of learning in a situation where the . organism must learn to avoid poisonous foods after a single exposure is an extreme example taste odor aversion conditioning Although potentiated of rapid acquisition , most odor conditioning appears to be acquired gradually over repeated trials. This form of associative conditioning may of higher learning in the human cortex. serve as an important mechanism Another essentialprinciple which appearsto have originally appearedin the olfactory system is parallel distributed processing (Kauer , 1991). The mammalianolfactory epithelium containssensorycells eachof which hasone of about one thousandgenetically different odor receptors(Axel , 1995). The axons from these sensory neurons project into the olfactory bulb to terminate on the approximately two thousand glomeruli, with primary olfactory neuronsexpressinga given receptor terminating predominantly on the same glomeruli (Axel , 1995). However, environmental scentsstimulate numerous , with eachodor causinga specificolfactory receptorswith different strengths and : Skarda Kauer 1994 1991 , , ; , ) ( Shepherd temporally (Freeman& spatially Laurent of 1995 1996 1985; Cinelli, Ha milton, and Kauer , ; , ) unique pattern activity in the olfactory bulb which is broadly distributed. Thus, the olfactory circuitry converts an environmental chemicalstimulus through a broad range of receptorsinto a complex pattern of activity in a large neuronal network which is capable of recognizing approximately ten thousand scents , Kienker, and (Axel , 1995). This pattern of parallel organization (Sejnowski distribution of and broad 1985 1995 , , ; , ) (Freeman activity Shepherd Shepherd 1987) serves as a template that is adapted by the cortex of the vertebrate mammalianbrain.

~Networks in Primates : Functional Ashford : Fuster Co~ tiVE , Coburn .&

TELENCEPHALON

MYELENCEPHALON
Figure 2.1 Fundamental components of the brain- telencephalon , diencephalon , mesen , and metencephalon . For the brain stem , and myelencephalon cephalon , ventral is left and dorsal is right . The cerebellumis not shown.

VERTEBRATE BRAIN ORGANIZA nON


Certain principles of vertebrate brain organization have been established , such as sensory analysesoccurring dorsally, motor direction occupying a ventral position, and autonomic function lying in an intermediate position. Also, segmentation developed , so that local sensation led to local motor activation. A later development specializedthe anterior segmentsfor more , Martinez, Shimamura , et al., 1994). In the vertebrate complex analysis(Rubenstein - the , the anterior five segments telencephalon(most anterior), dien, mesencephalon , metencephalon , and myelencephalon cephalon develop into the brain (figure 2.1), while the posterior segmentsbecome the spinal cord. In the higher vertebrate brain there is a further specializationfor sensory information analysis . The dorsal myelencephalonis specializedfor somato event detection (nucleuscuneatusfor upper limbs and nucleusgracisensory lis for lower limbs) and the dorsal mesencephalon is specializedfor auditory (inferior colliculus ) and visual (superior colliculus) event detection. These structures receive information through large, rapid transmission fibers and, therefore , serve as sentinels to analyze the sudden occurrenceof change in the environment. In contrast , the more anterior diencephalon (thalamus ) receives information from these modalities along direct, separate slower , . pathways for fine detail analysis

Introdudion to Neural Networks

gracilis

nucleus

pons

cuneatus

nucleus

METENCEPHALON

colliculus

colliculus

ENCEPHALON superior inferior

DIENCEPHALON thalamus . us hypothalam

Movement is regulated by severalstructures , including the metencephalic cerebellum , the ventral red nucleus and substantia nigra of the mesen and the basal of the . The cerebellum , cephalon ganglia telencephalon generates fast ramp movements , while the mes encephalicnuclei and the basal ganglia pace slow ramp movements (Komhuber, 1974). Accordingly , the brain divides motor activity functionally into fast ballistic movements and slow deliberateactions . the vertebrate brain, autonomic function continues to be Throughout regulated intermediately between dorsal sensory systems and the ventrally connectedmotor systems . In the autonomic nervous system , several brain levels coordinate cardiopulmonary function, temperature regulation, and . The anterior apex of the autonomic system is the hypothalamusin the sleep ventral diencephalon . The hypothalamus is largely responsible for coordinating , thirst, territoriality , and reproduction complex drives such as appetite . The hypothalamusis controlled in , and for fear and stressreactions , the &ontolimbic loop ( Nauta , 1971), and other telepart by the amygdala . A particularly important issuefor the autonomic system ncephalicstructures is the conservationof energy, an issuerelating to a variety of factors including , sleep , predator/ prey status , strategiesfor reproduction, ecological niche and brain size (Berger 1975 Allison & : Cicchetti , ; , 1976; Arm strong, 1983). - visual, auditory, and somatosensory - have The other sensory systems developed pathways into the cortex to take advantage of the information: Karten, 1970; Freeman & : processing power of this structure (Nauta & Skarda . 1995 This invasion has also , 1985; Karten, 1991; Shepherd , ) brought other neurotransmitter systemsinto the telencephalonto playa role in activation and information processing , including acetylcholine and GABA neurons es &om serotonin, norepinephrine , and projecting axonal process , and neurons whose cell in bodies lie mes , , and dopamine diencephalic encephalic ' metencephalicstructures 2.2 . (6gure ) THE ROLE OF THE CORTEX IN INFORMADON PROCFSSING The medial temporal lobe structures in primates are considered nontopo: Bower, 1989; Kauer , 1991; Axel , 1995; graphically organized (Haberly & , 1995). These regions have no direct input &om somatosensory , Shepherd or visual but do receive &om the brain , , auditory systems activating inputs stem and diencephalon . In mammals , the lateral telencephalondeveloped a specializedstructure with six lamina referred to as neocortex (Killackey, 1995), the principal structure in the primate brain for processingcomplex information. As other sensory systemshave invaded the cortex, primary regions with specialized topographic organization have developed (somatotopic organization for somatic sensation , cochleotopic organization for audition, and retinotopic for vision ). As the sensory systems establishedtheir primary organization

Ashford, Coburn : Fuster : FunctionalCognitive Networks in Primates . &

NUCLEUS BASALIS OF MEYNERT (cholinergic)

, elaboration of the sensory processing entry regions behind the central sulcus cerebral volume regions pushed developmentposteriorly . As the sensory also established close relationshipswith the medial , they systemsdeveloped temporal lobe structures for the evaluation of the importance of sensory information to the well-being of the animal (amygdala ) and spatial categorization (hippocampus ) of information (figure 2.3). Somatomotor function invaded the neocortex just anterior to the central sulcus and in conjunction with the somatosensoryregion, which formed just posterior to this sulcus . , the primary motor cortex has a somatotopic organization Consequently which is closely coordinated with the primary somatosensoryregion. The somatomotor cortex establisheda close relationship with the basal ganglia , putamen (caudate , globus pallidus) for pacing and directing movements(figure 2.4). Elaboration of motoric activity for vocalization (Preuss , 1995), and and 1974 cortical volume , ( ), presumablythought planning Matthysse pushed in with the cortex , development anteriorly primates prefrontal coordinating with the nucleus accumb ens for pacing speechand abstract thought . Thus, the neocortex of mammalsplays a role in all sensoryand motor function, the telencephalonexpanding over the lower brain regions both anteriorly and the increased . posteriorly to accommodate processingdemands An important and long-standing controversy has addressedthe question of the organization of information processing beyond the primary cortical regions. Though topographic organization has developed several levels of

Introduction to Neural Networks

RAPHE

serotonergic (

MIDBRAIN

dopaminergic (

NIGRA

SUBSTANTIA

campus

hippo

to

FORNIX (

LOCUSCOERULEUS (NORADRENERGIC )

CENTRAL

PRIMARY PARIETAL LOBE FORNIX

AMYGDALA

TEMPO PARIE JUNCT


OCCIPITAL CORTEX

LOBE
. The temporal , parietal , and , and memorysystems , perception sensory Figure 2.3 Posterior with the and are in bidirectional communication information sensory occipitallobesprocess . These alsoprojectto andamygdala lobe medialtemporal , includingthe hippo regions campus this thanthey less dominant in their influence on structure but are the basal , ganglia probably . than the frontal lobe is on the basal lobe structures or areon themedial ganglia temporal

CAUDATE
PREFRONTAL

PRIMARY MOTOR CORTEX CENTRAL SULCUS

CORTEX

NUCLEUS

PUTAMEN
GLOBUS PALLIDUS
-, and thought-coordinating systems . The frontal cortex Figure 2.4 Anterior motor -, speech constitutes the in the nucleus accumb ens which into the basal , , heavily ganglia particular projects large anterior portion of the basalganglia. However, the frontal lobe seemsto have less direct . influenceon the medial temporal lobe structures

: Fuf\dional Cognitive Networks in Primates Ashford, Coburn I: Fuster ,&

: Van complexity in primary and secondaryneocortical regions (Felleman& Essen 1992 , 1991 ; Van Essen areas of , Anderson, and Felleman the , ), large neocortex still seemto lack such organization , even as they have expanded to meet the processingdemandsof complex environmental niches (Lashley , 1950). For example , the temporal lobe has pushed anteriorly in primates to meet the need for more elaborate analysis of visual information (Allman, 1990). Yet the anterior temporal lobe has no significant retinotopic organization : Gross , 1979; Tanaka (Desimone& , Saito, Fukada , et al., 1991; Tanaka , 1993; Nakamura : Kubota, 1994). , Mikami , & Important considerationsfor understandinginformation processingin the brain are timing and coordination. The primary thalamicnuclei relay detailed information to the primary sensoryregions of the cortex. However, relevant broad cortical associationregions are activated synchronously with the primary -detecting neurons of the brain regions, presumably by the occurrence stem acting through the pulvinar of the thalamusor by the reticular activating : Magoun, 1949), which includes ascending monosystem (Moruzzi & aminergic and cholinergic pathways and the reticular nuclei of the thalamus : Everitt, 1995). Also, some modulation of input may occur (Robbins & " efferent " control (Pribram , 1967). Cortical activation in responseto through a stimulus is evidencedby electrical field potentials recordableat the scalp . Following cortical activation and receipt of detailed information, analysisof stimulus particulars occurs in the cortex with reciprocal communication occurring between all of the activated cortical regions (for reviews, see : Fuster , Szwarcbart , Mishkin, et al., 1965; Ashford & , 1985; Coburn, Kuypers Ashford, & : Fuster , 1990; Ungerleider , 1995). PRIMATE CORn CAL SENSORY , PERCEPTUAL , AND MEMORY

SYSTEMS
Visual System
Many of the inferencesregarding neuropsychological information processing in the human brain are derived from studies of the monkey. The most . In primates , there is a widely studied models involve the visual system of retinal hemifields to both the contralateral unique crossing superior colli culus and the primary visual cortex (Allman, 1982). Primary visual cortex is activated retinotopically by photic stimuli, and neurons are found there which preferentially respond to bars of light with unique orientations. These neurons are organized in slabs alternately serving inputs from the left and : Wiesel, 1977). The monkey cortex contains at least right eyes (Hubel & twenty additional visual areassurrounding the primary visual cortex which are responsiblefor analyzing a variety of discrete aspectsof visual information . Injury to a discrete areacan causeloss of a specificneuropsychological analysis capability. The areasmost closely connected to the primary visual cortex have a high degree of retinotopic organization, which diminishes at

Introduction to Neural Networks

Figure 2.5 Infonnation transmissionbetween different regions of the brain. The dorsal and ventral pathways leading forward from the occipital cortex are shown connecting all the way to . Short and long fibers connect the sensory regions across the specific frontal cortical regions central sulcus . The auditory region' s connections with the temporal lobe are shown . Each of theseregions has many other connectionswhich are not shown.

higher organizational stageswithin the secondaryvisual areas(Felleman& Van Essen , 1991). Beyond the primary and secondary visual areas , retinotopic influence on neuronal responsesbecomesdifficult to detect (Desimone & Gross , 1979 ; Nakamuraet al., 1994). The alternative considerations are whether the specific pattern of analysis is yet to be determined or the mode of distributed processingprovided by the olfactory model is utilized. As processingproceedsforward &om the primary and secondary visual areas , information is processed along two separate functional pathways (Ungerleider & Mishkin , 1982) (figure 2.5). One pathway leads ventrally toward the inferior temporal lobe. This ventral pathway abstractssuchvisual details as color, shape , and texture for identification of objects (Kuypers et al., 1965). In monkeys , a specializedregion in the posterior inferior temporal , Albright , region seemsto playa role in the analysisof faces(e.g., Desimone Gross , et al., 1984; Mikami , Nakamura , & Kubota, 1994; Ungerleider, 1995), though the intensity of neuron responseto facesin this region may simply indicate the general importance of face analysis , even in the monkey (Desimone it is unclear whether an association , 1991). So even in this unusual case , is . Farther in forward the inferior temporal cortex, region specialized neuronsrespond to many stimuli (Desimone& Gross , 1979). At the anterior of the lobe neurons , tip temporal respondpredominantly to abstractstimulus Nakamura et al. 1994 without , ( ), any clear evidence of topographic aspects whether classificational , , or otherwise. An important organization retinotopic, question regarding the nature of neuron responses along this path &om primary visual cortex to the tip of the inferior temporal lobe concerns the selectivity of individual neurons for specificenvironmental items or characteristics . In the primary and secondary regions, individual neurons show a broad range of responsesbetween high selectivity and nonselectivity ( Van

Ashford : Fuster : Functional in Primates , Coburn ,& CO2l \itjve Networks

VISUAL

PRIMARY

"

'

) '

. .

'

"

;;

i :

~ -

~ ~ ~ ~ E ; : ~ : : ~ : ~ : ~ : . . : : . : ~ ; ~ ~ :

PRIMARY MOTOR

PRIMARY SOMA : TOSENSORV PRIMARY AUDITORY

Essen& Deyoe, 1995). Inferior temporal neurons also have certain degrees of stimulus selectivity, but most neurons can readily be found to respond to one member of any limited set of stimuli, and neurons rarely show highly exclusive selectivity (Tanaka , Saito , Fukada , et al., 1991; Tanaka , 1993; Nakamura , Mikami, & Kubota, 1992; Nakamura et al., 1994). The range of selectivity in the inferior temporal cortex suggests that a stimulus which activatesa neuronal field will elicit responses from many neuronsrather than a few unique neurons , implying a broadly distributed analysis of information , a pattern of stimulus representationanalogousto that of the olfactory . system These findings concerning visual perception in monkeys are relevant to humans . However, in humans , there is clear evidenceof hemisphericspeciali zation. The left hemisphereis usually specializedby encoding verbal information . Regardingrecognition of faces , Milner (1974) found that in humans , lobe lesions interfere with the to remember faces and right temporal ability irregular line drawings, but did not affect memory for (perhaps easily verbally encoded . ) geometric shapes The secondvisual pathway leads dorsally from the secondaryvisual cortex toward the parietal cortex and is responsiblefor analysis of spatial relationships . In the dorsal pathway, spatial analysis of visual information is performed in conjunction with posteriorly projecting connections from the ' somatosensorycortex which monitors the animals own position. Both animals and humansshow deficits in learning tasks requiring perception of the , following lesions of the posterior parietal cortex. In humans , body in space body image and perception of spatial relationshipsare often severely abnormal following parietal injury . In addition to the two specific visual pathways describedabove , projecting from the retina to the primary visual cortex and then anteriorly, neurons at all levels of the visual cortex receive activating input from the pulvinar (e.g., Benevento& Rezak , 1976; Macko, Sarvis , Kennedy , et al., 1982). The visual information from a retinal , pulvinar receiving rapid projection through the superior colliculus , activates the visual cortex broadly, priming neurons at all levels of both visual pathways to analyzeinformational details arriving through the retinal geniculostriatepathway. Neurons of the inferior temporal visual cortex are sensitive to behavioral state , including attention (Maunsell, 1995). Neurons in this region have a substantial background level of activity , respond to stimuli with approximately the same latency as the neurons of the primary visual cortex, and remain elevated in the level of activity for severalhundred milliseconds following visual stimulation (Ashford & Fuster , 1985). Further, they respond to stimuli as a differentially presented repetition after less than two intervening stimuli (Baylis & Rolls, 1987). While neurons in this region can be classifiedto some extent according to the range of objects to which they , Rolls, & Leonard , 1987), the selectivity of different neurespond (Bayliss rons' responses to a wide variety of stimuli can vary consider ably (Nakamura

Introduction to Neural Networks

et al., 1992, 1994). Nearly half of the neurons in the inferior temporal region will respond to one of two simple visual stimuli in the context of abehav Loralparadigm which requiresattention to eachstimulus when it is presented : Fuster (Ashford & , 1985; Coburn et al., 1990) (figure 2.6). This suggeststhe existenceof an extensive functional neural network (ensemble ) comprised of , within which analysis of the roughly half the inferior temporal neurons . The 50 percent responserate is a behaviorally relevant stimulus takes place level which mathematicallyallows the most powerful analysis of any stimulus , 1972; Coburn et al., 1990). The lower limit of the responserate (John would be one neuron responding to a single environmental configuration, requiring a unique neuron for eachconfiguration. Clearly, this situation is an inadequateexplanation and even a small number of responding cells could not provide adequateinformation-processingpower to account for information : Richmond , 1993; Singer , 1995b ). To achievemaximal encoding (Gawne & encoding capability, the optimal responselevel is 50 percent of neurons in a field being activated by an environmental stimulus. Higher proportions would give lesspower, as do smallerproportions. Approx ~ mating a transient 50 percent responserate would also allow cortical modulating process es to ensure maximal distribution of processing across cortical regions, while maintaining stability of neuronal excitation (figure 2.7). Further gradation of neuronal responses , for example by modulation of response frequency , would provide additional analytic power. An important issue in the mode of information analysis of environmental events in the cortex is the nature of the temporal sequencingof the analytic es. Early anatomical investigations suggested that information processing process was serial , following the hierarchy from primary to secondary to associationregions of the cortex. The presumption was that processing at each level took some finite amount of time before the results of that processing could be relayed to the next higher level. However, the discovery of : Pandya , reciprocal anatomicalconnections(Kuypers et al. 1965; Rockland & 1979), simultaneously supported the concept of efferent control along the visual cortical hierarchy (Pribram , 1967). Thus, it became apparent that information processing could involve reciprocal communication along the identified processing pathways , even as far as the medial temporal lobe Mishkin & : . 1981 When concurrent processingwas discoveredat ( ) Aggleton , the initial and terminal ends of the ventral visual cortical pathway (Ashford & : Fuster , 1985) (seefigure 2.6), and as far as the hippocampus(Coburn et al., 1990), the notion of simultaneoushierarchical processing was introduced, , including suggesting that quite complex methods of analysis were possible . parallel distributed processing Individual neuronal responses are organized into temporally brief bundles : Fuster (Ashford & , 1985), which have a statistical distribution (Bair, Koch, Newsome , et al., 1994). Additionally , some information may be encoded in the temporal structure of the spike trains (McClurkin , Optican, Richmond , et al., 1991; Eskandar : Optican, 1992; Ferster & : Spruston , Richmond ,& , 1995;

Ashford, Coburn, & Fuster : FunctionalCognitive Networks in Primates

: Figure 2.6 Monkey performing delayed rnatch to- sample task (Ashford, 1984; Ashford & stimulus a Rash from the after a 20 second . In this Fuster 1985 , task . , ) upper panel waiting period ' illuminates the monkey s visual field. Exactly 2.0 secondslater, the top button of the triangle is illuminated either red or green. The monkey must not touch any of the three stimulus buttons for at least 0.5 second prior to the Rash , until the top stimulus light is illuminated. Then the the trial to continue. The light is darkenedafter 1.5 seconds button for the must monkey top press , then the monkey must wait 10 secondsfor the two lower lights of the triangle to be illuminated , either red and green or green and red. The monkey must push the button whose color matchesthe sampleto get a juice reward. Then the waiting period begins again. Shown below recorded from either the occipital or inferior are compositesof severalneuronal unit responses . Note that the occipital button was illuminated red in which the stimulus for a trial cortex temporal cortex units respond to the Rashwhile the inferior temporal units are largely inhibited. However, units from both the occipital and inferior temporal cortex respond to the samplestimulus . Vertical dashedlines separate0.5-secondepochs , and the , and over a similar time course small hashmarks represent20 ms.

Introduction to Neural Networks

'

'

SAMPLE

OCCIPITAL CORTEX

. I . . I . .

'

'

. I I . I I I . .

. I I FLASH . I I INFERO TEMPORAL CORTEX

8/16 /16 , p=12 ,870 15 , p=16


Possibilities N = number of cells M = activated cells

. . = = / / 1 6 1 1 62 120 16 ,p ,p 1111

rate .A Figure 2.'1 Matrix showingthe mathematical power of the so percentresponse darkened circlecould represent a responding a neuron , while an opencirclecould represent neuron . For N if one half of N, the number of , very nonresponding large Misapproximately is aboutIN, possibilities

: Gray, 1995), or spatiotemporalfiring patterns (Abeles, Prut, Bergman Singer & et al., 1994; Singer , , Engel , Kreiter, et al., 1997). However, reciprocal information transfer forward and backward across numerous cortical processing stages(seefigure 2.5) is probably required for memory storage (Rolls & : Treves , 1994), and can include coherent neuronal activity occurring across distributed sites (Bressler : Nakamura , Coppola, & , 1993; Bressler , widely 1995). Independentof potentially complex temporal responsepatterns of cortical neuron ensembles , the massivelyparallel anatomicalarchitectureof the cortical system provides great power for storing and recognizing imagesusing a vector convolution and correlation approach (Murdock, 1982). In this construct ' , eachneuron s responserepresentsa component of the vector occupying a huge N-dimensionalabstract mathematicalspace , with N representing the number of neurons in the brain. In this model, the total number of potentially encodable environmental configurations approximates 2N (see figure 2.7; for large N , if M approximates! x N , the number of possibilities es 2N), a satisfactorily large number. Recentstudieshave supported approach the concept that memories of details about the various attributes of a discrete visual object are stored in a distributed fashion in the respectivemultiple regions responsible for the sensory analysis and perception of those specific attributes (Ungerleider, 1995). This approach for storing information at the neuronal level can be viewed as a vector convolution operation Murdock , 1982), using the NMDA receptor (McClelland et al., 1986) or ( other long-term potentiating mechanisms , and involving the establishment of new connections between different neuronal systems (Alkon , 1989), as well as the altering of the efficiency of existing synaptic connections to

Ashford, Coburn I: Fuster : FunctionalCognitive Networks in Primates ,&

change their weighting . Recognition occurs if there is a significant correlation between the vector of a perceivedimage and the vector which describes . the current state of the cerebralsystem The implications of this model of reciprocal , hierarchicalinformation processing can be examined in the monkey in the visual processing pathway. Early electricalactivity in the cortex (as early as 20 ms in the monkey cortex in responseto a flash , but over 40 ms for a discrete visual detail) represents . Unit and field responsescan be modified by alertness initial visual processing : Vaughan , 1975) and attention to specific detail (e.g ., Arezzo, Pikoff, & : Fuster , 1985; Maunsell, 1995), as field potentials in the 200 ms (Ashford & . The initial neuronal responsesuggests latency neighborhood are in the human that information about the visual stimulus is carried both to the primary visual processingarea to begin detailed analysis , and more widely to the entire visual system where it servesan alerting function, preparing the larger system for the synchronousand reciprocal analysisof visual information between the primary, secondary , and associationalsensory processing areasand the medial temporal lobe. Electricalsignalscorrespondingto selective attention, the analysisof discrete stimulus features , and the detection of a variety of types of unexpectedevents can be recorded&om both primates and humans . For example , recognition of information as discordant , in the human &om expectation (Donchin, 1981) or containing details which are to i, Karis, & : Donchin, 1986) will generate a late positive be retained (Fabian electrical signal (P300 , positivity at 300 ms), which is likely to indicate that the cortex has perceived the incoming information and has initiated a storage i , Karis, & : Donchin, operation on the perceived information (Fabian 1990). While much of the work on gross electrical activity following environmental events has been done using electrical recordings &om scalp electrodes in the human , studies in monkeys have shown comparable gross also allowing microelectrode analysis of concurrent while electrical events , . Microelectrode analysishas shown, for local activity in deep brain structures , Zola,that the P300 is not just localized to the temporal lobe (Paller example , et al., 1988), further supporting the concept that information Morgan, Squire over broad cortical regions, probably using a parallel distributed occurs processing processingmode. Work with monkeys looking at the responsesof single neurons to behaviorally relevant stimulus dimensionsshows the relationship between individual neuronal activity and cognition and also gives direct evidencethat a single neuron can participate in a variety of functional networks (Fuster , these functional networks are widespread , 1995). Furthermore ) , 1988) and can adapt to task (i.e., environmental (Goldman Rakic : Rolls, 1987). demandsover short periods of time (Bayliss& With regard to detail memory function, two systemshave been identified in the monkey (Mishkin , 1982), one involving the hippocampusand Papez circuit through the anterior nucleusof the thalamusand the cingulate cortex , 1937), and the other involving the amygdala and the Nauta circuit (Papez through the dorsomedial nucleusof the thalamusand the orbito&ontal

Introduction to Neural Networks

cortex (Nauta, 1971). The hippocampus seems to involve place memory ' : Nadel, 1978), and recent studies suggest that hippocampal neurons (O Keefe & in the monkey code for specific geographic directions which can be associatedwith visual information for storage organization (O 'Mara, Rolls, Berthoz , et al., 1994; Ono, Nakamura , Nishijo, et al., 1993; Rolls, Robertson, -Francois & : Georges 1995 . In contrast , ) , the amygdala codes for emotions, : Everitt, 1997) including fear (in rodents; LeDoux, 1995; Killcross, Robbins ,& and alimentary , emotional (in dogs: Fonberg , 1969), sexual (in monkeys: : Steklis , 1976), and social factors (in monkeys : Brothers & : Ring, Kling & 1993), and these factors can also serve to index visual information for retrieval (LeDoux, 1994 , 1995). Classic studies of monkeys with lesions of the amygdala revealeddisruptions of social behavior (Kling & : Steklis , 1976; Pribram , 1961; Rolls, 1995) deriving from failures to perceive or retain social cuesrelating to dominanceor sexualhierarchies . The visual cortex connects with the : Pandya broadly , 1975a ,b; Van hippocampus (Van Hoesen & Hoesen : Mesulam , Rosene : Van Hoesen ,& , 1979; Rosene& , 1987), allowing the hippocampusto facilitate information storage throughout the visual cortex (Ungerleider , 1995). However, only the anterior portion of the temporal cortex connectswith the amygdala (Krettek & : Price , 1977; Turner, Mishkin , & : Knapp, 1980; Mishkin & : Aggleton , 1981), allowing the amygdala to focus on facilitation of the analysis , encoding , and retention of more abstractvisual information, particularly that information related to function such as food and sexual appeal and poison and danger signals . This model is consistent with electrical stimulation experiments with the human amygdala which evoke memoriesassociated with emotions (Penfield , 1958).
Auditory System

The auditory system of the monkey is considered to process information using principles akin to those of the visual system (see figure 2.5). While it has been more difficult to train monkeys to perform tasks in response to auditory information , neurons of the auditory cortex are particularly responsive to the vocalizations of other monkeys . Further , there are multimodal cells between the visual and auditory regions within the temporal cortex . Somatosensory System

The somatosensory systemanalyzesinformation in the parietal cortex, but in close associationwith the motor system and the frontal cortex anterior to the central sulcus(Pandya& : Kuypers, 1969; Jones& : Powell, 1970; Pandya& : Yeterian, 1985) (see figure 2.5). The parietal cortex shows neuronal responseswhen monkeys perform touch discrimination tasks that are comparable to visual discrimination tasks . In a haptic delayed match-to -sample task (a tactile discrimination task with a delay), neurons in rhesus primary somatosensorycortex (51) exhibit memory properties by firing during the

Ashford, Coburn, & r: Fuster : FundinnaJ Co2nitive Networks in Primates

delay. Also, units in monkey parietal association cortex discharge during : Fuster , 1992). perception and mnemonicretention of tactile features(Zhou & THE FRONTAL LOBE AND ATrENT I ON AND ACTIVE MEMORY

SYSTEMS
Historically, there has been considerableeffort to define the systemsof the frontal lobe, particularly with regard to attention, thought, and decision " " es. The working memory concept was introduced by Carmaking process 1935 ( ) to explain the effectsof principal sulcuslesions in monkeys lyle Jacobsen . lesions of the and was later shown to apply only to vi suospatial tasks , inferior prefrontal convexity interfere with delayed responsetasks , whether or not there is a spatial component , by decreasing the ability to inhibit incorrect responses . lesions of the arcuateconcavity leave delayed response behavior unaffected , but decreasethe ability to choose between specific when presentedwith specificstimuli. Orbitofrontallesions reduce responses . This last finding led Egas Moniz to emotionality and emotional arousal the treatment of psychiatric patients with prefrontal lobotomy . attempt Patients with prefrontal lobotomy show an inability to change response strategieson the Wisconsin Card Sorting test (Milner , 1974), which appears to be the sameinhibition deficit seenin monkeys with inferior frontal convexity . However, these patients show little diminishment on measures lesions . Thus, an important function of the frontal lobes appears of intelligence of various actions to be weighing the consequences , and selection of some actions with the inhibition of others, within the repertoire of all learned or . possiblebehaviors The frontal cortex managesinformation by coordinating activity in the . The frontal sensory and perceptual regions posterior to the central sulcus cortical regions form a functional executive network with the dorsomedial nucleus of the thalamus and the basal ganglia (Alexander, Delong , & Crutcher, 1992) to generate smooth motor acts (Kornhuber, 1974) and behavioral sequences , 1997b), which include thinking and planning (Fuster (Matthysse, 1974). The frontal cortex makes reciprocal connections with both the ventral and dorsal visual pathways as well as the auditory and somatosensorysystems(see figure 2.5). Neurons in the frontal regions are active during tasks requiring visual attention, particularly during the period when short-term retention of information is required for spanning a delay interval before a responsecan be produced (Fuster , 1973). The prefrontal cortex can selectively analyze face information (0 Scalaidhe , Wilson, & -Rakic Goldman and the detail and 1997 , ) integrate spatial information , Rainer , & Miller , analysisperformed by the posterior cortical regions (Rao responses 1997). Presumably , the prefrontal cortex servesto organize and sequence in posterior perceptualregions (Goldman-Rakic , 1988; Fuster , 1997b). In this fashion , the frontal cortex participatesin the attentive aspectsof perception " " and active short-term memory, also referred to as working memory

Introduction to Neural Networks

(Goldman-Rakic , 1995), as well as facilitating the encoding of relevant information and orchestrating retrieval from long-term storage (Fuster , 1995, 1997a ). Development of Monkey Tasks for Attention and Active Memory The early paradigmswhich were developed for testing behavior in the monkey used tasks such as delayed alternation. As an example of this task, the monkey might be required to push one button, then several secondslater, push a different button. Behavior in such tasks is impaired by lesions of the frontal lobes. Early explanationsof the cognitive requirementsforperforming this task focusedon memory. However, the performanceof this task and others like it are more dependenton attention, or active short-term memory, than the long-term storage of information (Fuster , 1997b ). In a modification of the delayed alternation task -to , the delayed match sampletask, the correct button choice dependson matching to a previously , it is clear thdt it is attention to displayed sample(seefigure 2.6). In this task detail, then maintaining in an active state an internal representationof that now-absent samplestimulus image , which is critical to correct performance . Further studiesof the brain regions involved in performanceof this task have shown that the inferior temporal lobe is required for the analysis of the information detail component of the task. However, performance of the match when a significant delay is introduced depends on the frontal lobes , indicating that the capability of maintaining attention over time to an internal mnemonic representationof the stimulus detail is critically dependenton frontal lobe function. This dichotomy also demonstrates the behavioral interaction between the temporal and frontal lobes (Fuster , 1997b).

Tasksto Distinguish Active Memory and Retentive Memory


A critical issue in the understanding of memory function was the development of tasks which would demonstrate the storage of information after distraction (beyond the limits of attention, active short-term memory, or " " working memory ). An important early demonstration by Gaffan (1977a ,b) showed that monkeys could perform recognition tasks involving complex , multiple colors, and multiple spatial positions. Gaffan also demonstrated pictures that the fornix (seefigure 2.2), presumablybecause of its critical ana tomical role in connecting the basal forebrain to the hippocampus , was critical to the function of retentive memory, a form of memory frequently impaired in humanamnesicpatients (Scoville & Milner , 1957). The role of the medial temporal lobe was further clarified by Mishkin -to -sampletask with trial unique objects. (1982) using the delayed non-match This task showed impairment from medial temporal lobe lesions or cholinergic inhibitors, which both produce pronounced deficits of long-term . For example , in the case of HiM ., who had surgical memory in humans

Ashford : Fuster : Functional , Coburn ,& in Primates CognitiveNetworks

ablation of both medial temporal lobes for epilepsy control, he is profoundly impaired in the ability to acquire most forms of new information, but can recall information learnedprior to the surgery, and he can learn new motor ' skills (Scoville & Milner , 1957). In Mishkin s initial experiments , lesions of either the hippocampusor amygdala impaired the performanceof this task, and lesions of both systemsrenderedthe monkey incapableof retaining the ' critical information. Later studiesby Mishkin s group suggestedthat the rhinal cortex, which is close to both the amygdala and hippocampusbut projects more widely , plays the most pivotal role in the retention of information chevalier , Mishkin , et al., 1993). Of note, the rhinal cortex (Meunier, Ba includes the entorhinal cortex, which is the region of the human brain that ' seemsto be the initial site of attack of the Alzheimer s process (Braak & Braak , 1991). More recently, efforts have been made to computer-automate memory tasks for monkeys so that a computer screen can deliver the stimuli, and can be registeredusing a joystick or touch-screentechnology. Use responses ' of a joystick keepsthe animals hand out of the visual field, which is important sincethe hand itself can representa visual stimulus. Computer control of -reward contingenciesallows teaching of consider ably more complex response . For example tasks , rhesusmonkeys can be taught to read and associate with individual letters of the alphabetand retain those associations responses for up to an hour (figure 2.8) (Ashford & Edwards , 1991). Accordingly , such tasks can be used to test reaction times and brain wave components , and . Also, memory taskscan be applied both are comparableto human measures

A
+-

B
~ +~

EFGHIJKL
- + +-

. Figure 2.8 Monkey reading letters and performing joystick responses(Ashford at Edwards 1991). The monkey could push the joystick to either the left or right . This task was taught to the . the letters A . B. C. D always monkey by reinforcing responseswith sweetenedjuice rewards required the same direction for reward. However. the remaining eight letters were associated with a different pattern of corred directions each day. One monkey could learn each of these letters after a single trial (reward would meancorred diredion ; no reward would mean that the other diredion was corred ). Then it could recall the corred pattern for the letters. often flaw. after up to a half-hour delay. interspersedwith trials of the other letters. lessly

Introductior1to Neural Networks

to other modalities : Mishkin , suchas the sensorimotor , system(Murray & Fuster . 1984 : ; Zhou & , 1992 ) RELA nON BETWEEN PROCFSSING CAP ABILITIESAND ENERGY

REQUIREMENTS
In the course of understanding the relations between the volume of the cerebral cortex, information-processing power, and the need forenergy providing nutrients, the focus of attention must be on the neuron with its dendritic and axonal process es. There are about 50,000 pyramidal neurons under each squaremillimeter of cortical surfacearea . Eachneuron may have to 100 000 a dendritic tree which may extend over , arranged along up inputs 6 mm in many directions. Its axon has a comparable number of outputs which may extend as far as the baseof the spinal cord, but commonly several centimetersto a target cortical region. The surfacearea of a pyramidal neuron may average 1 mm2 . Therefore , but may reach 1 cm2 , the total neuronal membranesurfacearea under a squaremillimeter of cortex surfaceis about 50,000 mm2 . This equals about 10,000 m2 for the 2000 cm2 of the humancortex. About 40 percent of the metabolismof the brain is devoted to maintaining the resting potential acrossthis huge amount of neuronal surface membrane . The remainder of brain metabolism is devoted to active cell : Pellerin , 1996), 10 percent for recovering from action ( activity Magi stretti & and 50 for potentials percent synaptic activity . The brain consumes20 percent of the energy resourcesof the body; maintaining cerebral function is clearly a major energy cost to the individual. If a single unique stimulus were coded by a single neuron, the brain would indeed be quiet, and the demands of repolarization and synaptic activity would be minimized. However, the stimulated neuronal assemblies , whether they are hard-wired primary cortical connectionsor plastic networks in association cortex, include large proportions of the cortical neurons . Observations of monkey cortical units responding to relevant stimuli suggest that a functional network of about half of the neurons in a field can manifest a response to a stimulus and that individual neurons may respond with several : Fuster , 1985; Coburn et al., 1990; dischargesto a single stimulus (Ashford & Mikami et al., 1994; Nakamuraet al., 1994). During intenseneuronal activity neuronsdepolarizefrequently, creating a major metabolic expensein repolarization demand . The large proportion of cerebralcortex that is activated by environmental stimuli demandsa heavy supply of energy, particularly in the primate (Arm strong, 1983). The physiological demands of processing in relation to cognition on a regional cortical basiscan be visualized clearly by techniqueswhich measurelocal cerebralblood flow (single photon emission computed tomography and functional magnetic resonance imaging) and metabolism(positron emissiontomography). However, when the brain does achieve its maximum processing power, it may also achieve its point of

Ashford : Fuster : FunctionalCognitive Networks in Primates . Coburn, &

optimal processing efficiency and actually minimize its metabolic demand (Parkset al., 1989). One important question is, How do neuronsin the brain stem regulate the activity of neuronal ensemblesin the cortex? ( Woolf , 1996). For example , acetylcholine neurons project to limited cortical patches as small as a few : Mesulam , 1984; Wainer & , 1990), presumably square millimeters (Saper them into action for relevant processingrequirements . Serotonin neurons calling es are the most widely distributed in the brain, presum , whose process ably activate broad regions of the cortex during a variety of waking behaviors (Jacobs& : Azmitia, 1992). In relation to neural network models , catecholamines may playa unique role describableas adjusting the gain of -Schreiber & : Cohen, logistic activation functions in the network (Servan 1992). Thus, the projections from the brain stem seem to play the role of efficiently orchestrating the processingof the distributed neural networks to assimilateand respondto the environment. DEVELOPMENT OF INFORMA nON -PROCESSINGCAPACITY In early development , primary cortical regions undergo critical periods when environmental input directs the formation of neuronal connections(Hubel & : Wiesel, 1977) and the development of functional assemblies , 1995b (Singer ). ) regions also pass through critical Higher-order association(e.g., perceptual of time in life Ba : Ungerleider, 1995). In , chevalier ( Webster ,& early periods one example neonatal , damageto the inferior temporal cortex of the monkey was nearly fully compensated , as measuredby multiple memory tasks in four- to five-year-old animals(Malkova, Mishkin , & : Bachevalier , 1995). Yet, while the primary regions becomerelatively hard-wired early in life, higherorder associationperceptualregions appear to retain some plasticity, forming new connections to accommodate the learning of new information , 1995a ) or until disabledby injury or throughout life (Diamond, 1988; Singer a neurodegenerative Ashford Shan Butler , , , et al., 1995). By contrast, process( the medial temporal region, which is more primitive in its development , does not seemto be able to recover from injury at either early or late stagesof life (Malkova et al., 1995). The frontal cortex seemsto function less than optimally in immature animals(as the adjective implies when referring to human childlike behavior), but this region undergoes a critical period in humans with massivechangesin connectionsin late adolescence and early adulthood (Feinberg , 1987). Subsequentto this, less flexibility (e.g ., for personality . ) is present change As applied to neural networks, most cortical regions seemto have an early quiescentperiod, followed by a critical period in which the network undergoes intense learning and revision of connections , followed by maturity, after which the region achievesa particular pattern of function that is less modifiable. However, some brain regions , such as the middle temporal lobe

Introduction to Neural Networks

and the inferior parietal lobe, may retain high levels of plasticity throughout ' the animals life and maintain maximum ability to store complex information. (This maintained elevated level of plasticity may predispose these brain ' : ; Ashford & regions to the pathological changesseen in Alzheimer s disease Zec, 1993; Ashford et al., 1995). It is the task of the whole brain working in concert to perceive the environment, analyze relevant information, store critical new information, and develop plans for the future which facilitate the survival and reproductive success of the organism in a complex world . The of is important aspect development the formation of connectivities and the coordination of processingwithin and between brain regions. FUTURE PRIMATE MODELS FOR NEUROPSYCHOLOGY AND

NEURALNETWORKS
A central theme of this chapter is the value of nonhumanprimates as models of human cognitive process es. Important information about human cognition has been obtained from human studies , such as brain imaging and recording . However, such studies are brain electrical activity from scalp electrodes limited in the amount they can tell us about the structural substrates of information processing . When functional principles and more details are . In needed , valuable information can be obtained from animal studies humans , there is a constant challengeto study ever smaller and deepercomponents of neuronal networks. Animal researchextends the inquiry to progressively more basic levels. Studiesof animalsprovide meaningful answers to questions about human brain structure and function. Moreover , comparative analysiscan reveal cluesto the development of functions. The approach of studying the brain of a monkey performing a cognitive . Monkeys appear task will continue to be a valuable model for neuroscience to enjoy playing simple video games for extended periods of time, as do humans , and those gamescan be designed to test the capabilitiesof the animals . Brain function can be monitored using minimally invasive electrode . Perturbations can include administration recordings or scanning techniques of drugs with reversibleeffectsor transient lesionssuchas those achievedby es , the role of specificbrain structures , chemical cooling. Using suchapproach can be and , systems neuropsychological functioning explored relatively . noninvasively in nonhumanprimates An important future direction is the better understandingof how so many neurons work together within specific brain regions and across many different brain regions. Implantation of multiple indwelling electrode arrays , which can be monitored in concert with imaging procedures , and massive , computer analysis of the interactions between the individual neurons , and complex behavior, will reveal more information regional activations . The development of about how the brain functions in health and disease neural network and massiveparallel distributed processingmodels basedon empirical data obtained from across the primate order perhaps will reveal

: FunctionalCognitive Networks in Primates Ashford. Coburn. & : Fuster

insights into human brain function that will ~ anscend models developed , lower animals , or humansalone. using computers

REFERENCES
Abeles : Vaadia , M ., Prut, Y., Bergman , E. (1994). Synchronizationin neuronal transmission , H., & and its importancefor information processing . Progress in Brain Research , 102 , 395- 404. Alexander : Crutcher , G.E., Delong , M .D. (1992). Do cortical and basal ganglia motor , M .R., & " areasuse " motor programs to control movement ? Behavioral Brain Research , 15 , 656- 665. Alkon , D. L. (1989). Memory storageand neural systems . Scien , July, 42- 50. HficAmerican Allison , T., & : Cicchetti, D. V . (1976). Sleepin mammals : Ecologicaland constitutional correlates . Science , 194 , 732- 734. Allman, J. (1982). Reconstructingthe evolution of the brain in primatesthrough the use of comparative data. In E. Arm strong : D. Flak (Eds .), Primate neurophysiologicaland neuroanatomical , & brain evolu Hon : Mtthods and concepts . (pp. 13- 28). New York: PlenumPress Allman, J. (1990). Evolution of neocortex . In E. G. Jones : A . Peters(Eds .), Cerebral cortex ,& , Vol . 8A (pp. 269- 283). New York: PlenumPress . Arezzo, J., Pikoff, A ., & : Vaughan , H. G. (1975) The sourcesand intracerebral distribution of , 90, 57- 73. auditory evoked potentials in the alert rhesusmonkey. Brain Research Armstrong, E. (1983). Relativebrain size and metabolismin mammals . Science , 220 , 1302- 1304. Ashford, J. W. (1984). CorHcal in the monkey . PhiD. dissertation . of visualencoding electrophysiology Ann Arbor , MI : University Miao6lms International. Ashford, J. W., & : Edwards , H. (1991). Memory in the monkey: An automatedparadigm and the effect of nicotine. Society Abstracts , 17 , 1236. for Nt Uroscience Ashford, J. W., & : Fuster , J. M . (1985). Occipital and inferotemporalresponses to visual signalsin the monkey. Erperimental , 90, 444- 466. Neurology Ashford, J. W., Shan : Schmitt, F. A . (1995). Temporal quantification , M ., ButlerS ., Rajasekar , A ., & " " of Alzheimer 's disease Ha , 6, 269- 280. severity: Time index model. Demen Ashford, J. W., & : lec , R. F. (1993). Phannacological treatment in Alzheimer' s disease . In ' R . Parks . lec, & : R. Wilson (Eds .), Nt Uropsychology ,R s disease and other demen Has of Alzheimer . (pp. 589- 614). New York: Oxford University Press . (1995). The molecularlogic of smell. Scien Axel, R , October, 154- 159. HficAmerican : Kandel , E. R. (1995). Molecular and structural mechanisms Bailey, C. H., & underlying long-tenn neurosciences memory. In MS . Gazzaniga(Ed.), 11It cognitive (pp. 19- 36). Cambridge , MA : MIT Press . Bair, W., Koch : Britten , C., Newsome , W., & , K. (1994). Power spectrumanalysisof bursting cells in areaMT in the behaving monkey. Journalof Nt Uroscience , 14 , 2870- 2892. : Rolls, E. T. (1987 of neuronsin the inferior temporal cortex in short ) . Responses Baylis, G. C., & tenn and serial recognition memory tasks . Experimental Brain Research , 65, 614- 622. : Leonard , C. M . (1987 Baylis, G. C., Rolls, E. T., & ) . Functionalsubdivisionsof the temporal lobe neocortex . Journalof Nt Uroscience , 7, 330- 342. Benevento : Rezak , L. A ., & , M . (1976). The cortical projections of the inferior pulvinar and adjacent lateral pulvinar in the rhesusmonkey (Macaca mulaHa ): An autoradiographic study. Brain Research , 108 , 1- 24.

Introduction to Neural Networks

. J. (1975). Bioenergeticfunctions of sleepand activity rhythms and their possiblerele,R Berger Rh ' on Proceedings vanceto aging. Feder , 34, 97- 102. -Rattoni, F., Coburn Bennudez : Garda , K. L., Fernandez , J., Chavez , A . F., & , J. (1987 ) . Potentiation of odor by taste and odor aversionsin rats are regulatedby cholinergic activity of dorsal hippo. Ph Rrm R tmd BthRvior , Biochemistry , 26, 553- 559. Cology campus : Braak . ActR Braak , H., & . E. (1991). Neuropathological staging of Alzheimer-related changes 82 239259. , , Rthoiogy Neurop -scale cortical networks and cognition. BrRin Rese Bressler Rrch Reviews , S. L. (1995 , 20, ). Large 288- 304. ., & : Nakamura Bressler , S. L., Coppola, R , R. (1993). Episodicmultiregional cortical coherenceat . NRfure , 366 , 153- 156. during visual task perfonnance multiple frequencies Brothers : Ring , L., & monkey with responses , B. (1993). Mesial temporal neuronsin the macaque R I BrRinRese Rrch selectivefor aspectsof sodal stimuli. BthRviovr , 57, 53- 61. : Kauer Cinelli, A . R., Hamilton, K. A , & , J. S. (1995). Salamander olfactory bulb neuronal activity observed by video rate, voltage-sensitive dye imaging. III. Spatial and temporal properties of evoked by odorant stimulation. ] oumRlof Neurophysiology , 73 , 2053- 2070. responses : Fuster enciesin temporal lobe , J. M . (1990). Visual responsela~ Coburn , K. L., Ashford, J. W., & RlNeuroscience structuresas a function of stimulus infonnation load. BthRvior , 104 , 62- 73. : Rusiniak Coburn , K. L , Garda, J., Kiefer, S. W., & . K. W. (1984). Taste potentiation of poisoned I Neuroscience R odor by temporal contiguity . BthRvior , 98, 813- 819. -selectivecells in the temporal cortex of monkeys . ] oUmRl Desimone , R. (1991). Face of Cognitive

Neuroscience , 3, 1- 8.

-selective : Bruce of Desimone , T. D., Gross , C. G., & , C. (1984 , R., Albright ). Stimulus properties in themacaque . Journal . inferiortemporal neurons , 4, 2051 2062 of Neuroscience in the temporal . Brain ., & cortexof the macaque Desimone : Gross ,R , C. G. (1979 ). Visualareas - 380 Research . , 178 , 363 Diamond . NewYork: TheFree Press . , M. C. (1988 ). Enriching heredity - 513 ? ... Surprise ! Psychophysiology . Donchin , E. (1981 , 18 , 492 ). Surprise ., & : aptican neurons in Eskandar , E. N., Richmond , B. J , L. M. (1992 ). Roleof inferiortemporal of information aboutvisualimages visualmemory , I. Temporal , recalled , and encoding images - 1295 . Journal . behavioral context , 68 , 1277 of Neurophysiology . iMKaris , D., & : Donchin Fabian , E. (1986 ). P300andrecallin an incidental memory paradigm - 308 . , 23 , 298 Psychophysiology of mnemonic in a Von iMKaris , D., & : Donchin Fabian , E. (1990 ). Effects strategy manipulation and 75 22 35. Restorff . Clinical , , Neurophysiology Electroencephalography paradigm - 508 . I. . Adolescence and mental illness . Science 1987 , 507 , 236 ( ) Feinberg , Fe : Van hierarchical in the primate cerebral Ueman , D. J., & Essen , D. C. (1991 ). Distributed processing cortex . Cerebral Cortex . ,1 , 1- 15 . Science Ferster : Spruston the neuronal code - 757 . , D., & , N. (1995 , 270 , 756 ). Cracking andamygdala in food intake motivation , alimentary ). Theroleof thehypothalamus Fonberg , E. (1969 - 358 andemotional reactions . ActaBiologica . , 29 , 335 Experimentia Freeman of chaotic EEG with a dynamic modelof theolfactory , w . J. (1987 ). Simulation patterns - 150 . . Biological , 56 , 139 Cybernetics system

: Fuster : FunctionalCognitive Networks in Primates Ashford, Coburn, &

Freeman : Skarda , W. J., & , C. A . (1985). Spatial EEG patterns , non-linear dynamics and perception : The neo - Sherringtonianview. Bmin Research Reviews , 10 , 147- 175. -responseperfonnance Fuster , J. M . (1973). Unit activity in prefrontal cortex during delayed : Neuronal correlatesof transient memory. Journalof Neurophysiology , 36, 61- 78. Fuster , J. M . (1995). Memory in the cerebral cortex : An empiricalapproach to neuralnetworksin the humanand nonhuman . Cambridge . , MA : MIT Press primate Fuster , J. M . (1997a in Nenrosciences ). Network memory. Trends , 20, 451- 459. Fuster : Anatomy , J. M . (1997b ). Theprefrontalcortu and neuropsychology , physiology of thefrontal lobe . , 3rd ed. New York: RavenPress ' GaHan , D. (1977a ). Monkeys recognition memory for complex pictures and the effect of fornix transection . QuarterlyJournalof Erperimental , 29, 505- 514. Psychology GaHan , D. (1977b ). Recognition memory after short retention intervals in fornix -transeded . QuarterlyJournalof Erperimental monkeys , 29, 577- 588. Psychology Gawne : Richmond , T. J., & , B. J. (1993). How independentare the messages carried by adjacent inferior temporal cortical neurons ? Journalof Neuroscience , 13 , 2758- 2771. -Rakic Goldman , P. S. (1988). Topography of cognition: Paralleldistributed networks in primate associationcortex. Annual Reviewof Neuroscience , 11 , 137- 156. -Rakic P. S. (1995). Cellular basisof Goldman , 14 , 477- 485. working memory. Neuron : Wise, R . A . (1988). Comparisonsand refractory periods for medial forebrain bundle Gratton , A ., & fibers subservingstimulation-inducedfeeding and brain stimulation reward: a psychophysical , 438 , 256- 263. study. Brain Research

- 264 ? Trends in Neuroscience . , 12 , 258 memory

: Bower, J. M . (1989). Olfactory cortex: Model circuit for study of associative Haberly, L. B., &

, F. A. (1952 Order : An inquiryinto thefoundations ). TheSensory Hayek . of theoretical psychology : University of Chicago Press . Chicago Hubel : Wiesel , D. H., & , T. N. (1977 architecture of macaque ). Functional visualcortex . monkey B, 181 , Series , 1- 59. of theRoyal Proceedings Society of London : Azmitia , B. L., & and functionof the brain serotoninsystem Jacobs , E. C. (1992 ). Structure . - 229 Reviews . , 72 , 165 Physiological of frontalassociation , C. F. (1935 Jacobsen areas in primates ). Functions . Archives of Neurology - 569 andPsychiatry . , 33 , 558 size andtheevolution , H. J. (1991 Jerison . NewYork: American ). Brain Museum of Natural of mind . History versus , E. R. (1912 statistical theories John of learning ). Switchboard andmemory . Science , 177 , 850 - 864 . . G. (1990 in thedevelopment ,E Jones andevolutionof primate ). Modulatoryevents neocortex . - 362 In E. G. Jones andA Peters .), Cerebral cortex (Eds , Vol. 8A (pp. 311 : Plenum ). New York . Press : Powell , E. G., & Jones , T. P. S. (1910 ) An anatomical study of converging sensory pathways - 820 within thecerebral cortexof the monkey . Brain . , 93 , 193 Kaba : Nakanishi , H., Hayashi , Y., Higuchi , T., & of anolfactorymemory , S. (1994 ). Induction by - 264 theactivation of a metabotropic . Science . , 265 , 262 glutamate receptor " Karten , H. J. (1991 ). Homologyandevolutionary ." BrainBehavior and originsof the neocortex - 212 Evolution . , 38 , 264

Introduction to Neural Networks

to odor codingin the andparallel of topography Kauer , J. S. (1991 ). Contributions processing 14 85. in 79 Trends Neuroscience . , vertebrate , olfactory pathway . In MS . : A neuroanatomical brain of the human H. P. 1995 , ( ). Evolution perspective Killackey Press . MA: MIT 1253 . . 1243 neurosdences , The . ) ( (Ed ), Cambridge cognitive pp Gazzaniga behavior -conditioned : Everitt Killcross , B. J. (1997 ). Differenttypesof fear , T. W., & , S., Robbins - 380 . . Nature i within amygdala nuclei , 388 , 377 mediated by separate in nonhuman for affiliativebehavior : Steklis , H. D. (1976 ). A neuralsubstrate primates Kling , A., & . - 238 andEvolution . BrainBehavior , 13 , 216 to their : An introduction cortex , andbasal Komhuber , cerebellum , H. H. (1974 ). Cerebral ganglia . Thirdstudyprogram .), Theneurosdences : F. G. Worden(Eds . In F. O. Schmitt& motorfunctions - 280 . , MA: TheMIT Press ). Cambridge (pp. 267 cortex to the cerebral fromthe amygdaloid . L. (1977 : Price ,J ). Projections Krettek , J. E., & complex - 722 . . Journal , 687 in the rat andcat , 172 andthalamus Neurology of Comparative : Rosvold , H. E. (1965 ). Occipitotemporal , M., & , M. K., Mishkin , H. G. J. M., Szwarcbart Kuypers - 262 . 11 . in therhesus , , connections corticocortical Neurology 245 monkeyErperimental . Symposium , 4, of the engram , Sodety . (1950 Biology for Erperimental , K. S ). In search Lashley - 482 . 454 - 476 . . Neuron andtunes , 16 , 473 Laurent , G. (1996 ). Odor images American . Scientific andthebrain , 50- 57. , June le Doux , memory , JE . (1994 ). Emotion fromfearto emotion in thebrain : Leaping of anemotional le Doux , JE . (1995 ). In search system - 1061 neurosdences .), Thecognitive ). Cambridge . In MS . Gazzaniga (pp. 1049 andconsciousness (Ed . : MIT Press , MA : Mishkin , M. , L., & . M., Sokoloff Macko , M., Shinohara , C, Miyaoka , K., Jarvis , C. D., Kennedy . 281 394 396 . Science with 2C14 visual , , the . ) 1982 ( ) Mapping primate deoxyglucose ( system : Relevance of brainenergymetabolism : Pellerin , L. (1996 ). Cellularmechanisms , P. J., & Magistretti . Annals disorders brainimagingandto neurodegenerative to functional of theNewYork . , 380- 1387 , 777 of Sciences Academy -termeffects tem neonatal of selective chevalier : Ba , J. (1995 ). Long , M., & Malkova . L. Mishkin . 212 226 109 Neurosdence . Behavioral in , and , lesions on learning memory monkeys porallobe transmission to dopamine : Relationships , motorcontroland , S. (1974 ). Schizophrenia Matthysse .), Theneurosciences , Thirdstudy & : F. G. Worden(Eds . In F. O. Schmitt extraction feature program - 737 . , MA: TheMIT Press ). Cambridge (pp. 733 ' of visualtargetsin cerebral . H. R. (1995 Maunsell ). The brains visualworld: Representation ,J . . Science cortex , 270 , 764 769 distributed . (1986 : the PDPResearch ). Parallel McClelland , D. E., & processing , J. L., Rumelhart Group models . and 2: Vol. in the microstructure , : Erplorations Psychological biological of cognition . Press MA: MIT , Cambridge : Gawne , T. J. (1991 ). Concurrent , B. J., & , L. M., Richmond McClurkin , J. W., Optican processing . Science in visualperception , 253 , neuronal encoded and complexityof temporally messages - 677 . 675 on visualrecognition : Murray chevalier , E. A (1993 ). Effects Meunier , M., & , J., Mishkin , M., Ba . cortexin rhesus andperirhinal ablations of the entorhinal andseparate monkeys of combined - 5432 . , 13 , 5418 Journal of Neurosdence in the to photographs : Kubota Mikami , K. (1994 ). Neuronalresponses , K., & , A., Nakamura . BrainResearch . Behavioural , 60 , 1- 13 of therhesus sulcus monkey temporal superior

: FunctionalCognitive Networks in Primates : Fuster Ashford, Coburn, &

Milner , B. (1974 : Scope andlimits ). Hemispheric . In F. O. Sdunitt : F. G. Worden ,& specialization - 89). Cambridge .), The neurosciences . Thirdstudy (Eds . (pp. 75 , MA: MIT Press program Mishkin , M. (1982 . Philosophical ). A memorysystemin the monkey Transactions of the Royal . Series B. 298 , 85- 95. Society of London Mishkin : Aggleton , M., & , J. (1981 contributions of the amygdala ). Multiple functional in the -Ari (Ed - 420 . In Y. Ben .), The : Elsevier monkey . (pp. 409 ). Amsterdam amygdaloid complex Moruzzi : Magoun , G., & , H. W. (1949 formation andactivation ). Brainstemreticular of the EEG . - 473 andClinical Journal . of Electroencephalography ,1 , 455 Neurophysiology Murdock . B. B. (1982 andretrievalof item andassociative ). A theoryfor the storage information . Psychological Review - 626. , 89 , 609 : Mishkin , E. A., & tactualas well as visualmemorydeficitsfollow , M. (1984 ). Severe Murray combined removalof the amygdala and hippo in monkeys . Journal , 4, campus of Neuroscience - 2580 2565 . Nakamura . , Matsumoto ,K : Kubota , K., Mikami , A., & , K. (1994 of ). Visualresponse properties in the temporal . Journal singleneurons , 71 , 1206 poleof behaving monkeys of Neurophysiology 1221 . Nakamura : Kubota , K., Mikami , A., & , K. (1992 in the monkey ). Activity of singleneurons of a visual discrimination task . Journal amygdala during performance , 67 , of Neurophysiology - 1463 1447 . Nauta , W. J. H. (1971 of the frontallobe : A reinterpretation ). Theproblem . Journal of Psychiatric - 187 Research . , 8, 167 Nauta : Karten , W. J. H., & , H. J. (1970 ). A general brain , with sidelights profileof the vertebrate on theancestry of thecerebral cortex . In F. O. Schmitt : F. G. Worden(Eds .), Theneurosciences ,& . Thirdstudy : Rockefeller Press . (pp. 7- 26). NewYork program University O'Keefe : Nadel , J., & as a cognitive , L (1978 . New York : Oxford University ). Thehippo campus map Press .
' O Mara, S. M ., RoDs : Kesner . P. (1994). Neurons responding to whole, E. T., Berthoz , A ., & ,R . Journalof Neurosdtnce body motion in the primate hippocampus , 14 , 6511- 6523. Ono, T., Nakamura , K., Nishijo, H., et al. (1993). Monkey hippocampalneuronsrelated to spatial and nonspatialfunctions. Journalof Neurophysiology , 70 , 1516- 1529. -Rakic 0 Scalaidhe : Goldman , S. P., Wilson, F. A . W., & , P. S. (1991). Areal segregationof face , 278 , 1135- 1138. processingneuronsin prefrontal cortex. Science Paller , K. A ., lola -MorganS ., Squire ., & : Hi Dyard , L R , S. A (1988). P3-like brain waves in

nonnalmonkeys and in monkeys with mMiai temporal lesions . BehavioralNeuroscitn Ct , 102 , - 725 714 . : Kuypers , D. N., & , H. (1969). Cortico- cortical connedions in the rhesusmonkey. Bra;n Pandya - 36. Research , 13 , 13 : Yeterian , D. N., & and coMedions of corticalassociation , E. H. (1985 Pandya ). Architecture areas . In A. Peters : E. G. Jones .), Cerebral corfu ,& : Plenum (Eds , 4, (pp. 3- 61). New York Press . , J. W. (1937 mechanism of emotion . In R . L. Isaacson ). A proposed .), Basic in Papez (Ed readings . (pp. 87- 109 ). Hunting neuropsychology ton, NY: Krieger Parks . W., Crockett ,R , D. J., Tuokko , H., Beattie . L., Zec , B. L., Ashford . F., , J. W., Coburn ,K ,R " " Becker : McGeer . G. (1989 , R. E., McGeer , P. L., & ,E ). Neuropsychological systems efficiency andpositron emission . I 270 282 . , , Journal of Neuropsychiatry tomography

Introduction to Neural Networks

Parks . W., Levine ,R , D. S., Long , D. J., Dalton , I. E., Weingartner , H., Fedio , P., , D. L., Crockett Coburn : Becker . (1992 , K. L., Siler , G., Matthews distributed , J. R., & , R. E ). Parallel processing and neuropsychology : A neuralnetworkmodelof Wisconsin cardsortingandverbal8uency . - 233 Review . , 3, 213 Neuropsychology Penfield in temporal localization anddeepsylvianareas , W. (1958 . Research Publishing ). Functional Association in Research Nervous and Mental Disease - 226 . , 36 , 210 for Preuss fromanimals to humans in cognitiveneuroscience , T. M. (1995 . In MS . ). Theargument - 1241 .), The neurosciences . (Ed (pp. 1227 ). Cambridge , MA: MIT Press cognitive Gazzaniga Pribram . In DE . Sheer(Ed .), Electrical , K. H. (1961 stimulation ). Limbicsystem of the brain - 320 : University of Texas Press . (pp. 311 ). Austin Pribram , K. H. (1967 . ). The limbic systems , efferentcontrolof neuralinhibitionand behavior - 336 in Brain Research . , 27 , 318 Progress Rao : Miller, E. K. (1997 of whatandwherein the primate , S. C., Rainer , G., & ). Integration prefrontal - 824 cortex . Science . , 276 , 821 Robbins : Everitt andattention . In Gazzaniga , T. W., & . (Ed , B. J. (1995 .), ). Arousal , MS systems - 720 The neurosciences . , MA: TheMIT Press (pp. 703 ). Cambridge cognitive Rockland ., & : Pandya of corticalconnections , K. S , D. N. (1979 ). Laminar origins and terminations of the occipital lobein the rhesus . BrainResearch , 179 , 3- 20. monkey Rolls . T. (1995 andconsciousness to understanding ,E ). A theoryof emotion , andits application - 1106 theneural basis of emotion . In MS . Gazzaniga .), The neurosciences (Ed (pp. 1091 ). cognitive . , MA: TheMIT Press Cambridge Rolls . T., Robert : Georges -Francois of space ,E son , R. G., & in the , P. (1995 ). Therepresentation . Society Abstracts . , 21 , 1494 for Neuroscience primate hippo campus Rolls : Treves in the braininvolvedin memoryandrecall , E. T., & . , A. (1994 ). Neuralnetworks - 341 in BrainResearch . , 102 , 335 Progress Rosene : VanHoesen formation of the primate , D. L., & brain :A , G. W. (1987 ). Thehippocampal reviewof somecomparative of and connedions . In E. G. Jones :A ,& aspects cytoarchitecture - 356 Peters .), Cerebral corfu : Plenum . Press , Vol. 6 (pp. 345 (Eds ). New York Rubenstein : Puelles , J. L. R., Martinez , S., Shimamura vertebrate , K., & , L. (1994 ). Theembryonic - 580 forebrain : Theprosomeric model . Science . , 266 , 578 of cerebral cortical afferent in therat. II. Magnocellular , C. B. (1984 ). Organization systems Saper - 342 basal nucleus . Journal . , 222 , 313 of Comparative Neurology Scoville : Milner afterbilateral , W. B., & lesions . , B. (1957 ). Lossof recent memory hippocampal - 21. andPsychiatry , Neurosurgery , 20 , 11 Journal of Neurology : Shepherd , T. J., Kienker models of , P. K., & , G. M. (1985 ). Simple Sejnowski patternrecognition discrimination . Society Abstracts . , 11 , 970 olfactory for Neuroscience of molecular . , G. M. (1994 neuron ). Discrimination Shepherd by the olfactoryreceptor signals - 790 Neuron . , 13 , 771 basis for sensory . In MS . Gazzaniga , G. M. (1995 ). Towarda molecular Shepherd perception - 118 .), The neurosciences . (Ed , MA: MIT Press (pp. 105 ). Cambridge cognitive -Schreiber Servan : Cohen networkmodelof catecholamine , D., & , J. D. (1992 , March ). A neural - 130 modulation of behavior . Psychiatric Annals . , 22 , 125 in neocortical : A hypothesis . In Gazzaniga , W. (1995a ). Timeascodingspace , Singer processing - 104 MS . (Ed .), The neurosciences . (pp. 91 , MA: MIT Press ). Cambridge cognitive

Ashford, Coburn : Fuster : FunctionalCognitive Networks in Primates , &

andplastidtyof corticalprocessing , W. (1995b ). Development architectures . Science Singer , 270 , 758 - 764 .
: Gray, C. M . (1995). Visual feature integration and the temporal correlation , W., & Singer . Annual Reviewof Neuroscience , 18 , 555- 586. hypothesis , W., Engel , A . K., Kreiter, A . D., Munk , M . H. J., Neuenschwander : Roelfsema . , S., & Singer , P. R : necessity (1997). Neuronal assemblies in CognitiveSciences , signatureand detedability . Trends , 1, 252- 261. : Moriya , M . (1991). Coding visual imagesof objects in the Tanaka . K., Saito , H. A ., Fukada , Y., & , 66, 170- 188. inferotemporalcortex of the macaque monkey. Journalof Neurophysiology of object recognition. Science Tanaka , K. (1993). Neural mechanisms , 262 , 685- 688. : Shepherd , G. M . (1993). Synaptic transmissionand modulation in the olfactory Trombley, P. Q ., & bulb. CurrentOpinionin Neurobiology , 3, 540- 547. Turner, B. H., Mishkin, M ., & : Knapp , M . (1980). Organization of the amygdalopetalprojections from modality-specificcortical associationareasin the monkey. Journalof Comparative , Neurology 1991 , 515- 543. studi~s of cortical mechanisms brainimaging , L. G. (1995 for memory. ). Functional Ungerleider - 775 Science . , 270 , 769 : Mishkin, M . (1982). Two cortical visual areas , L. G., & . In D. J. Ingle, R . J. W. Ungerleider Mansfield : M . A . Goodale (Eds .), Analysis of visual behavior(pp. 549- 586). Cambridge ,& , MA: MIT Press . Van Essen : FeUeman , D. C., Anderson , C. H., & , D. J. (1992). Information processingin the primate visual system : An integrated systemsperspective . Science , 255 , 419- 423. Van Essen : Deyoe, E. A . (1995). Concurrent processingin primate visual cortex. In M . , D. C., & S. Gazzaniga(Ed .), Thecognitive neurosciences . (pp. 383- 400). Cambridge , MA : MIT Press Van Hoesen : Pandya , G. W., & , D. N. (1975a ). Someconnectionsof the entorhinal (area 28) and . Brain Research perirhinal (area35) cortices of the rhesusmonkey. I. Temporal lobe afferents , 95, 1- 24. Van Hoesen : Pandya , G. W., & , D. N. (1975b ). Someconnectionsof the entorhinal (area 28) and . Brain Research , 95, perirhinal(area 35) cortices of the rhesus monkey. 1lI. Efferent connections 39- 59. Van Hoesen : Mesulam , G. W., Rosene , D. L., & , M . M . (1979). Subicular input from temporal cortex in the rhesusmonkey. Science , 205, 608- 610. Wainer, B. H., and Mesulam MM . (1990). Ascending cholinergic pathways in the rat brain. In M . Steriade& : D. Biesold ( Eds .), Brain cholinergic (pp. 65- 119). New York: Oxford University systems . Press Webster chevalier : Ungerleider , M . J., Ba , J., & , L. G. (1995). Transient subcorticalconnedions of inferior temporal areasTE and TED in infant macaquemonkeys . Journalof Comparative Neurology , 353 , 213- 226. Woolf , N. J. (1996). Global and serial neurons form a hierarchicallyarrangedinterface proposed to underlie memory and cognition. Neuroscience , 74 , 625- 651. ' Zhou, Y., & : Fuster , J. M . (1992). Unit dischargein monkey s parietal cortex during perception and mnemonicretention of tactile features . Society Abstracts , 18 , 706. for Neuroscience

Introducti Or1 to Neural Networks

Attention

and Neural Networks

MichaelI. Posner andRajendra D. Badgaiyan

OVERVIEW
About a hundred years ago, William James(1890) characterizedattention as a distinct psychophysiological phenomenon . In The Principles , of Psychology " he described attention as taking possessionof by the mind, in clear and vivid form, of one out of what seemseveralsimultaneouslypossible objects or trains of thought. Focalization are of its , concentration of consciousness essence . It implies withdrawal from some things in order to deal effectively " ' with others (pp. 403- 404). Much of James s description of attention is still . appropriate , until recently there was Despite the familiarity of attention as a concept almost complete lack of understanding about the brain mechanismsunderlying attentional control. This is not so surprising if one considers the fact that an understanding of attention depends on understanding of brain physiology, its mechanismof information processing , and more significantly, and a assessment of cognitive processlike attention. objective understanding As our understanding of human brain physiology develops , cognitive functions like consciousness and attention are no longer regarded as philo sophical or theological entities. We have come to recognize them as independent , intimately related to brain networks. This psychological domains view has led to neurophysiological characterizationof cognition in terms of the real-time activation of brain areasinvolved in cognitive processing , using the tools employed for functional exploration of the human brain. In recent , use of positron emission tomography (PET ) and functional magnetic years resonance imaging (fMRI ) techniquesin conjunction with the scalp or depth event-related potentials (ERPs ) have allowed us to view the activity of brain areasin real time (Posner& : Raichle , 1994). Our current progress in this area involves the combined use of brain imaging methods and analysisof the cognitive operations involved in complex human tasks . Currently availablemethods basedon hemodynamicssuch as PET and fMRI lack capability for real-time analysis at the speedneeded for cognitive operations (tens to hundredsof milliseconds ) and there is only limited successin directly localizing sourcesof scalp potentials. However,

taken together , these methods have begun to allow us to observe the operation of networks in real time . These networks consist of anatomical areas that have been related to cognitive operations . The identification of these operations is itself a difficult part of the problem of defining brain networks . Current attempts to integrate information available in various fields to formulate a common strategy for understanding cognitive processes have led to the realization that the brain has an independent , anatomically identi fiable attentional system . Attention - specific areas of the brain are selectively activated when the organism is engaged in a task and they interact with the sensory , semantic, and motor systems . Because our understanding of the visual system is the best developed , studies of the attentional system have most often concerned visual attention . Attention , however , involves all modalities and higher -order networks of control . This chapter emphasizes visual attention , but we examine other attentional networks as well and seek common principles of operation . In this chapter we consider some of the work involving neuroimaging , ERP, and cognitive techniques in the area of attention in order to provide a background for the construct i~n of more detailed ' models that involve the brain s attention system .

ATTENTIONAL SYSTEM
Attention involves three major functions : orienting to sensory stimuli ; executive functions , including detection of target events and control of cognitive operations ; and maintaining the alert state. Though knowledge of the precise neural mechanisms responsible for these operations is still inconclusive , many of the concerned brain areas and networks have been identified . Thus , the orienting network for visual attention is believed to involve posterior structures such as the parietal lobe , pulvinar , and superior colliculus . The executive network is located more anteriorly , and includes frontal midline areas and basal ganglia , while the attentive alert state is maintained by the brain vigilance network operating through the norepinepherine system arising in the locus ceruleus (Posner 8t Petersen, 1990 ; Posner 8t Raichle, 1994 ).

Orienting Network
Anatomically , this network is located posteriorly , and for the visual system it is largely responsible for operations needed to bring attention to a location or object and to shift attention from an unattended to a target location . It serves to enhance information at the attended location in comparison with that that occurs at relatively unattended locations (Posner & Presti, 1987 ). In terms of cognitive theories , this network mediates disengagement , engagement , and amplification of the attentional target . Parietal Lobe The ability to disengage attention from an unattended location in favor of an attended location is a necessary component of the atten -

Introduction

to Neural Networks

tional system . In a study in which PET scanningof subjectswas done while were , increasedblood flow was observed they engagedin attentional tasks in the area of the pulvinar when subjectswere required to separatea target location &om surrounding noise (la Berge : Buchsbaum , 1995; LaBerge & , 1990). However, when attention was shifted &om one location to the other, the main increasein activity was in the right and left superior parietal lobe (Corbetta, Miezin, Shulman , et al., 1993). These studies suggest that attentional operationsare quite specializedwithin the overall network. The function of the parietal lobe was clarified by studying patients having lesions in this part of the cerebral cortex. It was found (Posner , 1988) that with unilateral lesions have no patients problem in orienting to the target , but when the target is presentedon the contralateral presentedon the lesioned side field, it is either neglected or attended following a long response time. The reaction time of these patients for contralateral field targets was two to three times longer than that of normal control subjects . However, before presenting the target, when a cue indicating possible location was given, the response of these patients was similar for ipsilateral and contralateral . These findings indicate that the poor performance target locations on the contralesionalside is not due to a sensorydeficit and that the role of the parietal lobe can be rather specificto a disengageoperation. Laboratory studies and clinical data suggest that there are differencesin attentional functions of the right and left parietal lobes. The right lobe shows increasedblood flow when the task requires attentional shifts either in the right or left visual field, where the left parietal lobe is involved only in right visual field shifts (Corbetta et al., 1993). Another difference between right and left-lesioned parietal lobe patients is in their ability to detect local or . When patients are shown a large figure consisting of smaller global features , patients with left parietal damage neglect small figures and report figures , whereasthose with right hemispheredamagereport small figures large figures and neglect the large ones (Robertson, Lamb : Knight , 1988). Moreover ,& , right -lesionedparietal patients are generally slow comparedwith those with left parietal lesions (Ladavas : Provinciali, 1989). Finally, left, Pesce ,& lesionedparietal patients appearto be influencedby object boundariesmuch more than right -lesionedparietal patients. Thesefindings help to explain the clinical observation that right hemispherelesions produce greater attention " deficit than those on the left and " neglect as a clinical syndrome is more : Van Den prevalent with right than with left hemispheredamage(Hellman & Abell, 1980; Mesulam , 1981).
Besides the parietal cortex , a midbrain structure , the Superior Colliculus superior colliculus is also involved in the process of attentional shifts (LaBerge, 1995 ). The superficial layer of the superior colliculus receives input from retinal ganglion cells. It has been estimated that about 10 percent of the ganglion cells project to this layer (Van Essen , Anderson , & Felleman, . 1992 ) The deep layer has extensive connections with the frontal eye field

Posner & Bad : AttentionandNeural Networks ~aiyan

: , Spencer , Hardy, et al., 1981), prefrontal cortex (Goldman & (Leichnetz Nauta, 1976), and posterior parietal lobe (Lynch, Greybriel, & : Lobeck 1985 , ). Stimulation of the cells of this layer producesmovementsof the eye and the head (Hikosaka & : Wurtz , 1983), and these cells show increasedfiring just to saccadic movement (Goldberg & : Wurtz , 1970). It has been suggested eye prior that the movement component of attention shifts involves a network within the superior colliculus and the disengagementand engagementcomponents involve networks in the posterior parietal lobe and pulvinar, respec : Raichle , 1994). There is direct evidence of involvement of tively (Posner& nuclei of the thalamus in visual attention in several PET studies pulvinar Corbetta Miezin al. et : Buchsbaum , , Dobmeyer, ( , 1991; LaBerge& , 1990).
Information &om the attended object or location is ampli Amplification fied in comparison with other locations or objects . This relative amplification allows preferential higher -order processing of the target by the brain . When a target is attended , activity of the brain areas which specialize in its processing are selectively enhanced. Thus , in a PET study , wherein the subjects were required to attend to different aspects of a stimulus - its motion , form , or color - different prestriate brain areas were activated in each condition (Corbetta et al., 1991 , 1993 ). For example , while attending to a moving object , significant increases in blood flow were found in a prestriate area of the mid temporal lobe . This area might be the human equivalent of the V5 area of the monkey , which contains cells with a strong selectivity for moving stimuli . Further , there was increased blood flow in the motor area when subjects attended to motor movement (Roland , 1984 ). In general , attention to sensory information appears to increase blood flow in the specific brain areas which process that event . These activations appear to give relative priority to the event to which one is attending . However , it is still not clear exactly how this amplification is accomplished . It may involve mainly reduction of unattended areas in some cases and enhancement of attended areas in other cases . At neural and cellular levels , there are theories (Van Essen : , Anderson , & Olshausen , 1994 ) suggesting that the amplification is relative and that it results &om suppression of information &om unattended objects and locations : Desimone , 1985 ). (Chelazzi , Miller , Duncan , et al., 1993 ; Moran & Thus , when monkeys were trained to attend to a particular location to respond to a color change, cells in area V 4, representing a nontarget location , stopped responding to an otherwise adequate stimulus (Moran & : Desimone , 1985 ). Similar suppression has been observed in inferior temporal cortex during a visual search task. Cells that initially fired strongly to a visual object suddenly stopped responding when the monkey attended to a different object at another location (Chelazzi et al., 1993 ). , particularly in visual Suppression of nontarget information makes sense tasks. In fact , early visual cells are stimulated to maximum even during passive viewing , so there is little room for target enhancement, but ample scope

Introduction to Neural Networks

for suppression of nontarget objects . As the information ascends, there is enhancement of activation for target and suppression for nontarget objects . Thus , neurons of the retina , the lateral geniculate nucleus, and area VI fire strongly to both attended and nonattended objects while the cells of extrastriate areas such as V 4 show greater activity for the attended object (Moran & Desimone , 1985 ). It has been proposed (Felleman & Van Essen , 1991 ) that visual cells have narrow fields and early receptive respond mechanically to all stimuli , but as one moves up , the receptive fields widen and cells become more sensitive to the attentional targets and react less effectively to the unattended stimuli . Executive Network

An anteriorly located attentional network is involved in attentional recruitment and it controls the brain areasthat perform complex cognitive tasks . It exerts general control over the areas involved in target detection and response . This network is more active when the task involves complex discrimination , Pardo , Janer , et al., 1990) and the degree of activation (Pardo on the number of , Petersen , Fox, et al., depends targets presented (Posner 1988). The network is also responsiblefor anticipation of the target location (Murtha, Chertkow, Dixon , et al., 1995). Areas involved in this network are the midprefrontal cortex, including the anterior cinguiate gyrus and supplementary motor area (Posner& : Petersen , 1990). It may also involve areasof the basalganglia which supply dopamineto the frontal lobe. Norman and Shall ice (1986) have presenteda cognitive model that specifies some of the operations involved in this supervisory function for attention . Norman and Shallice assumethat the first level of cognitive control " " " to which usesa " schema operatesvia a contention-scheduling mechanism coordinate well-learned behaviors and thoughts. The contention-scheduling mechanismis involved when the task demandsroutine selection . However, when the situation is novel or highly competitive and conflicting, a second " intervenes level of control mechanismcalled the " supervisory system . This additional inhibition or to the activation system provides target signal for better perception and discrimination.
Generate Uses Task It has been observed that in attentional tasks involving conflict conditions , there is activation of the anterior cingulate gyrus . One such condition is the generate uses task in which the subject is required to think of the proper use of a visual or auditory word . When the ERP evoked by this task was subtracted from that elicited by passive repeating of the words , three cortical areas of activation were observed . The first activation was located in the anterior cingulate , the second in a frontal area anterior ' ' to Broca s area, and the third in Wernicke s area (Petersen, Fox , Posner, et al., 1988 , 1989; Posner, Sandson, Dhgwan , et al., 1989 ). Since the second and third areas are classic language areas, the first one , that is, the cingulate ,

Posner & : Bad : AttentionandNeuralNetworks ~aiyan

should be related to attention . To further examine the supervisory function of the cingulate , subjects were allowed to practice the same uses for the words over and over again to make the task simpler and less competitive . According to the Norman -Shallice model , this procedure should not use supervisory control for execution and should be mediated by the contention scheduling mechanism. That was exactly what was observed in a PET study (Raichle, Fiez, Videen , et aI., 1994 ). When subjects practice the same list repeatedly , generating the same use for each word , the anterior cingulate and left lateral activations disappeared . Interestingly , these activations reappeared when anew , unpracticed list was given to the same subjects .

Circuitry In order for a brain area to perform a supervisory function it must influencewidely distributed parts of the brain where computations related to the task are performed (Posner& Raichle , 1994). Anatomical studies like brain that the anterior , many cingulate regions, has close contact suggest with many other cortical areas (Goldman-Rakic , 1988). The cingulate connectionsto the lateral frontal areasinvolved in word recognition, and the posterior parietal areasinvolved in orienting are particularly strong. Recently efforts have been made to trace the dynamics of these interactions , 1995 for reviews , 1992; Rugg & Coles by use of ERPs(seeNaatanen the . While it is difficult to determine of the ERP methods ) generator from a scalpdistribution of electricalactivity , if the generator is known from PET or fMRI studies it is much easier to evaluate whether the scalp distribution could come from that generator (seeHeinze , Mangun, Burchert , et al., 1994, for an exampleof this methodology). The algorithms to relate a generator to the scalp distribution (Scherg & Berg, 1995) work best when fewer generators are involved. Thus in complex tasks , it is important to use a subtraction . which isolatesonly a small number of brain areas One effort to do so is in localization of a scalp negativity that follows , Posner , & Tucker, 1994; Gehring, Gross , Coles , et making an error (Dehaene al., 1993). When subjectswere aware of making an error in speededtasks , in a the localized a showed key press they very strong negativity following area over the midfrontal scalp . Further analysis using the brain electrical source analysis (BESA ) algorithm (Scherg & Berg, 1995) showed this error negativity most likely camefrom the anterior cingulate gyrus. Errors can be either slips, incorrect execution of a motor program, mistakes , or selectionof an inappropriate intention. The Norman-Shallicemodel would predict supervisory system involvement in the recognition of an execution of an incorrect motor program, but not if the contention-schedulinghad selectedwhat was . Dehaeneet al., (1994) have shown thought to be an appropriate response that error negativity follows a slip when the person knows the error, but not if he or she is unawareof the error (Tucker, Liotti , Potts, et al., 1994). he generate uses minus repeat words task described above Studies of ~ ERPs have shown evidenceof an anterior cingulate activation starting using about 170 ms after the visual word (Snyder , et al., 1995). , Abdullaev, Posner

Introduction to Neural Networks

This activation was presumably related to some kind of focal attention . The cingulate activation stayed present and was joined after 50 ms by a left &ontal activation . Abdullaev and Posner ( 1997 ) took a further step in replicating the PET results. They had subjects generate uses for the same list several times. The left &ontal and cingulate activations tend to go away following practice , but the activations were restored when a new list was presented or when subjects were required to generate a new use for the practiced words .

Stroop Test Another attentional task involving competitive and conflict condition is the Stroop test (Stroop, 1935). In this test, the subject is required to namethe ink color of a printed word. The word can be congruent (ie ., the word red printed in a red color); neutral (a noncolor-related word , e.g ., top printed in red); or incongruent (the word blue printed in red). In one such study (Pardo et al., 1990), imaging data were obtained by subtracting the congruent condition from the incongruent. The anterior cinguiate was found to be active in three areason the right medial portion and one areain the left . In another study (Bench inferior sulcus , Frith, Grasby, et al., 1993), an increasein right anterior cingulate activation was observedduring both congruent and incongruent conditions, when compared to neutral. This finding suggeststhat the effort to name ink color and avoid reading the word can influenceboth congruent and incongruent conditions in the sameway. This of cinguiate experiment also demonstratedthat there is further enhancement the task is made increased and activation when the attentional demandsare more competitive by increasing the number of stimuli or the rate of presentation . This is to be expectedif the cinguiate is performing a supervisory function. In another study of the Stroop-like effect (George, Ketter, Parekh , et al., 1994), emotional words were used instead of color words. This time there , was increasedactivation in the left midcinguiate , the left anterior cingulate . Further, when the tasks were self-paced , there and the right midcingulate was left midcingulate activation, and the right anterior cinguiate was active . These results suggest that different parts of the in externally paced tasks . in different are involved computational tasks cinguiate Thus, there is evidence suggesting involvement of the anterior cinguiate . It is therefore not surprising that in tasks that require difficult decisions lesions of the frontal lobe often produce disorganized and incoherent behavior . Thus the patients having dysexecutive syndrome following closed head injury , stroke, or degenerativedisorders of frontal structures have difficulty in solving problems which need prior planning (Duncan , Burgess ,& Emslie , 1991a , b). , 1995; Shallice& Burgess Cerebralstrokesin the areaof the cinguiate gyrus causea condition called akinetic mutism. It has been observed that these patients lack initiation for , 1994). Work with cats and monkeys spontaneous behavior (Damasio , 1955a , b) also suggestssimilar involving lesions of the cinguiate (Kennard

: Attention and Neural Networks Posner& 1 : Badgaiyan

lack of initiation for voluntary behavior and movement. To study the supervisory function of the cingulate in a patient, the Stroop test was performed both before and after cingulotomy (Janer : Pardo ,& , 1991). Following cingulotomy , there was a significant deficit in performancein the congruent condition , while in the incongruent condition the performancewas comparable to that in normal control subjects . Sinceboth the congruent and incongruent conditions require a similar selective attentional mechanism , it is not clear why there was no deficit in the incongruent condition. One possibility is that the brain areas involved in executive attention may be widely distributed within the midfrontal and supporting subcorticalarea . Other evidence suggesting involvement of the cingulate in supervisory attentional function comesfrom study of schizophrenicpatients. Early positive symptoms of schizophreniainclude difficulty in attention toward the right visual field and trouble in performing tasks involving Stroop-like conflict conditions (Maruff & : Currie, 1995; Posneret al., 1989). Thesesymptoms suggest dysregulation of supervisory or executive attentional functions, indicating involvement of the anterior cingulate (Early, et al, 1989a ; b). Language -related problems in these also indicated that the deficit was in patients the anterior cingulate , which exercisescontrol over both language and spatial attention. Interestingly, in a recent postmortem study, cellular abnormal ities were found in the cingulate of a schizophrenicpatient (Benes , 1993), suggestingpossible involvement of the cingulate in the expressionof schizophrenic symptoms.
Network Vigilance

The brain network responsiblefor maintenance of the alert state is an important of attention becausealertnessis required for the execution component of attentional tasks . Thus, the vigilance network should be a part of the brain's attentional system . The vigilance network comprisesthe noradrener at the locus ceruleus of the midbrain and gic pathways originating terminating at various locations in the cerebralcortex (Harley, 1987). This network is significantly activated when the subjects are required to maintain an alert state , particularly in the foreperiod of a reaction time task and while waiting for an infrequent target (Posner& : Petersen , 1990). the state heart rate and scalpelectricalactivities are attenuated , During vigilance and the blood flow in the right frontal and right parietal lobes increases . It appearsthat the integrity of the right frontal lobe is essentialto the maintenanceof attentional alertness . Patients with lesions in the right lateral frontal lobe show poor performancein tasks requiring alertness , and interestingly, these patients do not have bradycardia during the vigilance state ( Wilkins : McCarthy, 1987). Right frontal activation has also , Shallice ,& been demonstratedin the auditory vigilance task (Cohen , Semple , Gross , et ai, 1988). One of the important functions of the vigilance network is its abil-

Introduction to Neural Networks

ity to produce widespread inhibition of cortical and autonomic activities , 1973). When a subject is warned to attend to a target, the (Kahneman . The reduction provides amplitude of backgroundcortical activity is reduced smaller background noise and helps in effective amplification of the target event. There is evidence to suggest that at the cellular level, norepinephrine is involved in the mechanismwhich reducesbackground and enhancestarget : Posner , 1994). Thus, when a sympathomi , Marrocco, & activity (Jackson metic drug, amphetamine , the resting blood flow of the brain , is administered while the subject waits for the target and increaseson arrival of decreases -adrenerqic agonist, the target. Similarly, administration of clonidine, an !X2 & : Geffen results in abnormal distractibility (Clark, Geffen , , 1989). Children having attention deficit hyperactivity disorder (ADHD ) have excessivedistractibility , particularly if the task requires sustained attention for a long In fact, normal and ADHD children are the same when the target . period occurrswithin 100 ms of the cue , but, when the delay is increasedto 800 ms, the performanceof ADHD children is consider ably poorer than normal and , 1991). It has this is particularly true for right visual field targets (Whitehead been suggestedthat in ADHD , there may be a failure of the vigilance mechanism , resulting in inadequateamplification for relevant signal, leading to increased , Potkin, et al., 1991). , Posner distractibility (Swanson

OFTHENETWORKS INTERACTION
. The three attentional networks describedabove are closely interconnected Cortical components of the orienting and executive attentional networks communicatewith eachother directly via corti co-cortical pathways and both of them receive input from the vigilance network through catecholaminergic . In addition, prefrontal and posterior projections from the locus ceruleus parietal cortex project to , and receive input from, the other brain regions , 1988). implicated in the control of visuospatial function (Goldman Rakic the subcortical with also communicate networks these of Cortical components in attentional are involved elementslike the superior colliculus , that : Posner mechanism(Jackson , 1994). In addition, the executive , Marrocco, & network can modulate the activities of the orienting network indirectly through the basal ganglia. The striatum, a portion of the basal ganglia, has projections from both the prefrontal cortex, which is a part of the executive network, and the parietal cortex, which is a part of the posterior network. These connectionsallow the striatum to relate spatial information generated et al., 1994). by the two networks (Jackson
Alerting and Orienting

Connectionsbetween the visual orienting network and the vigilance system are strong. The locus ceruleussends significant noradrenergic input to the

Networks : AttentionandNeural Posner& : Bad ~aiyan

on & Foote, 1986). Using parietal, pulvinar, and collicular systems ( Morris these connedions, the vigilance system tunes up the posterior system and, , enables it to interact with the through amplification of relevant objects information accumulatingin the object recognition systems . Because of this interaction, during the highly alert state , reaction times are faster but subjects make more anticipatory reactions and have high error rates (Posner , 1978).

BasalGanglia
A number of models have appearedsuggesting specific interactions among attentional networks in the control of orienting (Jackson et al., 1994; la Berge , -Schreiber 1995; la Berge & Brown, 1989; Cohen, Romero , Servan , et al., 1994). Thesemodels postulate a distributed neural system the units of which interact to produce the observed functions of amplification of the attended target. The basal ganglia have been considered to be particularly important in mediating the connection between executive and orienting networks. The structuresof the basalganglia are closely connectedto the other brain areas involved in the attentional process(Jacksonet al., 1994; la Berge , 1995). The basalganglia are united with the cortex through cortico-striatal thalamocort ical circuits. It has been suggestedthat there are at least five such pathways through the basal ganglia, each organized in parallel and innervating different regions of the thalamus and cortex (Alexander, Crutcher, & Delong , 1990). The output of the basal ganglia is inhibitory and GABAergic neurons of the medial globus pallidus (GPm) and substantia nigra (SNr) are tonically active. They hold the thalamusand other structuresin a state of tonic inhibition . This inhibition is modulated via two distinct pathways which are organized in opposition to one another and link the striatum to output neurons in the GPm-SNr complex. Under normal condition, these pathways are sensitively balancedby dopaminergicinfluence(Gerfen , 1992). Further, since the basalganglia are the sourceof dopamine input to the anterior cingulate , thesetwo structureshave a close relationship. lesions of the basal ganglia result in spatial neglect ( Villardita, Smirni, & , 1983) and it appears that the basal ganglia are important for Zappala . Thus, parkinsonianpatients have switching attention between different sets reduced ability to shift between task sets (Hayes , Davidson, Keele , et al., 1998). It has been suggested that in tasks involving visual attention, the basalganglia coordinate closely with the frontal areasto bring about suitable orientation (Posner & Dehaene , 1994). It serves the important function of cortical and subcortical uniting regions implicated in the processingof attentional information.

Introduction to Neural Networks

CONTROL COGMTIVE
In order for a brain area to perform a supervisory function it must influence widely distributed parts of the brain where computations related to the task are performed (Posner& Raichle , 1994). Anatomical studies suggestthat the anterior cingulate like brain , many regions, has close contact with many other cortical areas(Goldman Rakic , 1988). The cingulate connectionsto lateral frontal areasinvolved in word recognition and to posterior parietal areas involved in orienting are particularly strong. What is the cingulate activation actually doing? According to our analysis the cingulate is involved in producing the local amplification in neural activity that accompanies top-down selectionof items. It is easiestto understand this function in the domain of processingwords. It is well-known from cognitive studies that a target word is processedmore efficiently following a related prime word (Posner , 1978). A portion of this improvement occurs automatically becausethe prime word activates a pathway sharedwith the target. However, another portion of the activation is top-down because attention to the prime word leads the subject to expect a particular type of target. If the prime is maskedor of low validity , the improvement in the target will be mostly automatic ; however, if the prime word is of high validity or if subjectsare instructed to use the prime word to think of another category , top-down effectsdominate. If the target is ambiguous(e.g ., the word ) can lead to a single consciousinterpretation palm), the prime word (e.g., tree that fits with both the prime and target words (see Simpson , 1984, for a review). We believe that the cingulate is responsible for these top-down effects by providing a boost in activation to items associated with the . expectation Anatomically, we seethe cingulate in contact with areasof the left lateral and posterior cortex that seemto be involved in understandingthe meaning of a given target word. Indeed , the time courseof activation of the cingulate 170 ms and the left lateral frontal ( ) (220 ms) cortex found during the generate our that attention interacts with the semantic task support speculation activation pattern (Snyder et al., 1995). Prime words presentedto the right visual field produce rapid activation of close semanticassociates , whereasprime words presentedto the left visual field act more slowly and activate remote associates as well. Blocking attention to the visual word by shadowing (Nakagawa , 1991) reducessemantic and a that resembles that found when the priming produces pattern priming to the left word is visual field prime presented irrespective of the visual field . This finding suggestsa specificrole of to which the prime word is presented attention in producing the left hemispherepriming pattern. Similarly, prime words that are masked and those not subject to specific attentional effect basedon their meaning activate a wider range of associates than those that are unmasked . These findings show evidence of the specific role of supervisory . control over semanticactivation patterns

Posner & : Badgaiyan : AttentionandNeuralNetworks

In addition , we believe that cingulate activation plays a role in the voluntary reactivation of brain areas that can also be driven automatically from input . Feature analysis of visual targets appears to involve right lateralized posterior parts of the brain . If subjects are instructed to examine a feature : Raichle, voluntarily , similar electrode sites are activated much later (Posner & 1994 ). We believe that the cingulate is important to these activation patterns . We also believe that cingulate activation is involved when elements of a thought are reordered in time . By increasing activation of the brain area that performs a specific computation , one can change the time course of the : Raichle, 1994 ). organization of the component operations (Posner & The idea of attention as a network of anatomical areas makes relevant study of both the comparative anatomy of these areas and their development in infancy (Posner & : Raichle, 1994 ). In the first few months of life , infants adult abilities to orient to external events , but the cognitive develop nearly control produced by the executive attention network requires many months or years of development . Studies of orienting and motor control are beginning to lead to an understanding of this developmental process. As more about the maturational processes of the brain and transmitter systems is understood , it could be possible to match developing attentional abilities with changing biological mechanisms. The neural mechanisms of attention must support not only common development among infants in their regulatory abilities but also the obvious differences among infants in their rates of , and successin , attentional control . There are many disorders that are often supposed to involve attention , including neglect , closed-head injury , schizophrenia , and attention deficit : Raichle, 1994 ). The specification of attention hyperactivity disorder (Posner & in terms of anatomy and function might be useful in clarifying the underlying bases for these disorders . The development of theories of deficits might also foster the integration of psychiatric and higher -level neurological disorders , both of which might affect the brain ' s attentional system .

REFERENCES
Abdullaev, Y. G., & : Posner , M . I. (1997). Time course of activating brain areasin genera tin ~ verbal associations . Psychological Science , 8, 56- 59.

- thalamocortical Alexander : Delong , M . R. (1990). Basal , G. E., Crutcher , M . D., & circuits ganglia " " and"limbic : Parallel substrates for motor , oculomotor " functions . In H. B. M. , prefrontal , C. G. Van Eden , j . P. C. DeBruin : M. G. P. Feenstra .), Progress , M. A. Comer in ,& Uylings (Eds - 146 brainresearch : Elsevier , Vol. 85 (pp. 119 . ). NewYork

Bench , C. j ., Frith , C. D., Grasby , P. M., Friston : Dolan , K. j ., Paulesu . , E., Frackowiak , R. S. j ., & ,R of the functional j . (1993 ). Investigations . Neuro anatomyof attentionusing the Strooptest - 922 . , 31 , 907 psychologia Benes , F. M. (1993 of cingulate cortexto schizophrenia ). Relationship andotherpsychiatric disorders . In B. A. Vogt & : M. Gabriel(Eds .), Neurobiology corlex and limbicthalamus of cingulate - 605 : Birkhiuser . (pp. 581 ). Boston

Introduction to Neural Networks

Chelazzi : Desimone basis for visualsearch in , L., Miner , E. K , Duncan , J., & , R. (1993 ). A neural - 347 inferiortemporal cortex . Nature . , 363 , 345 : Geffen and covert orientation of Clark , C. R., Geffen , G. M, & , L. B. (1989 ). Catecholamines - 139 attention in humans . Neuropsychologia . , 27 (2), 131 -Schreiber Cohen : Farah of spatial , J., D., Romero , R. D., Servan , D., & , M. J. (1994 ). Mechanisms attention : the relationship of marcrostrudure to microstructure in parietalnegled . Journal of - 387 Neuroscience 6, 377 . Cognitive Cohen . Dowling : Nordahl , R. M., Semple , W. E., Gross , M., Holcomb , H. J , T. E. , s. M., & of sustained attention . Neuropsychology andBehavioral (1988 ). Fundionallocalization , Neurology I , 3- 20. Corbetta : Petersen . E. (1991 and , M., Miezin , F. M., Dobmeyer , 5., Shulman , G. L., & ,S ). Selective dividedattentionduringvisualdiscriminations of shape : Functional , color , and speed anatomy - 2402 . Journal emission . , II , 2383 by positron of Neuroscience tomography Corbetta : Petersen , M., Miezin , F. M., Shulman , G. L., & , S. E. (1993 ). A PETstudyof visuospatial attention . Journal Neuroscience 13 3 1202 . 1226 , ( ), of ' error Damasio : Emotion brain . New York : Putnam . , A. R. (1994 , reason , andthehuman ). Descartes Dehaene : TuckerD. M. (1994 of a neuralsystem for error , 5., Posner , M. I., & ). Localization - 305 detection andcompensation . PsychologiC RIScience . , 5(5), 303 Duncan : Emslie after frontal lobe lesion . Neuro , J., Burgess , P., & , H. (1995 ). Fluid intelligence . , 33 (3), 261 268 psychologia : Raichle of the left , T. 5., Posner , M. I., Reiman , E. M., & , M. E. (1989a ). Hyperactivity Early -pallidal striato . PartI: Lowerleveltheory . Psychiatric . , 2, 85- 108 Developments projection : Raichle , T. 5., Posner , M. I., Reiman , E. M., & , M. E. (1989b ). Hyperactivityof the left Early striata - pallidalprojection . Part II: Phenomenology and thought disorder . Psychiatric Developments . - 121 , 2, 109 : Van Essen in the primate hierarchical Felleman , D. J., & , D. C. (1991 ). Distributed processing cerebral cortex . Cerebral Cortex ,1 (1), 1- 47. : Donchin , B., Coles , M. G. H., Meyer , D. E., & , E. (1993 ). A neural system Gehring , W. J., Gross - 390 for errordetection andcompensation . Psychological Science . , 4, 385 ., Ketter , M. S , T. A., Parekh , P. I., Rosinsky , N., Ring , B. J., Trimble , M. R., George , H., Casey Horwitz : Post brain activity when selecting , B., Herscovitch , P., & , R. M. (1994 ). Regional : An Hls0 PETstudy of the Stroopand an emotionalStroop . response despiteinterference - 209 Human Brain . , 1(3), 194 Mapping mosaic : Multiple levelsof compartmental . Gerfen , C. R. (1992 ). The nigrostriatal organization - 139 Trends in Neurosciences . , 15 , 133 . H. (1970 : Wurtz of eye movement and stimulation on units in ,R ). Effects Goldberg , M. E., & colliculus . Federation . , 29 , 453 monkey Proceedings superior : Nauta . H. (1976 demonstration of a projection from Goldman , P. 5., & , W. J ). Autoradiographic association cortexto superior in the rhesus colliculus . BrainRese Rrch , 116 , monkey prefrontal - 149 145 . -Rakic Goldman of cognition : Parallel distributed in primate networks , P. S. (1988 ). Topography - 156 association cortex . Annual Review . , 11 , 137 of Neuroscience in arousal ? Progress in , C. W. (1987 , emotionand learning ). A role for norepinephrine Harley - 458 I Psychiatry andBioiogi C R . , 11 , 419 Neuropsychop Mnnacology

Posner& : Badgaiyan : AttentionandNeuralNetworks

: Rafal . D. (1998 , A. E., Davidson , M. C., Keele , S. W., & ,R Hayes ). Towarda functional analysis of thebasal . Neuroscience 10 178 198 . , Journal , of ganglia Cognitive : Van Den Abell Hellman , K. M., & , T. (1980 dominance for attention : The ). Right hemispheric - 230 mechanism of inattention . (neglect ). Neurology , 30 , 227 asymmetries underlying hemispheric Heinze . , Burchert , H. J., Mangun , G. R , W., Hinrichs , H., Scholtz , M., Muntel , T. F., Gosel , A., ., Hundeshagen : Hillyard , H., Gazzaniga , M. S., & , S. A (1994 Scherg , M ., JohannesS ). Combined attentionin humans . spatialand temporalimagingof brain activity during visual selective - 546 Nature . , 372 , 543 Hikosaka : Wurtz , 0 ., & , R. H. (1983 functions of monkeysubstantia ). Visualandocculomotor IV: Relationof substantia . Journal nigra pars reticulata nigra to superiorcolliculus of Neuro - 1301 . , 49 , 1285 physiology . R., Marrocco : Posner JacksonS , R., & , M. I. (1994 areas ). Networksof anatomical controlling - 944 visualspatial attention . Neural Networks . , 7(6/ 7), 925 . Vol. 1. New York , W. (1890 : Holt. James ). Principles of psychology . W., & : Pardo . V. (1991 in selective ,K ,J attention Janer anterior dn). Deficits followingbilateral - 241 . Journal Neuroscience . , 3(3), 231 of Cognitive guiotomy -Hall. andeffort . Englewood : Prentice Kahneman , D. (1973 Cliffs ). Attention , NJ Kennard , M. A. (1955a . Journal Jous and ). Thecinguiate gyrusin relationto consciousness of Ntn MentalDiseases , 121 (1), 34- 39. ablation of the dnguiate Kennard , M. A. (1955b area on the behavior ). Theeffectof bilateral of cats . Journal . - 169 , 18 (2), 159 of Neurophysiology 's art mind la BergeD. (1995 : The brain . Cambridge ). Attentional , MA: Harvard processing of fulness Press . University la BergeD., & : Brown in shape , V. (1989 identification . Psychological ) Theoryof attentional operations - 124 Review . , 96, 101 la BergeD., & : Buchsbaum emission measurements of pul, M. S. (1990 ). Positron tomographic - 619 vinaractivityduringanattention task . Journal . , 10 , 613 of Neuroscience Ladavas : Provincial i , L. (1989 attentiondeficits , E., Pesce , M. D., & andhemispheric ). Unilateral - 356 in the controlof visualattention . Neuropsychologia . , 27 , 353 asymmetries Leichnetz . , Spencer : Astruc , G. R , R. F., Hardy corticotectal , S. G., & , J. (1981 ). The prefrontal in the : An and horse radish . Neuroscience monkey projection anterograde retrograde study paroxidase - 1041 . , 6, 1023 : Lobeck , J. C., Greybriel , AM ., & , L. J. (1985 Lynch ). The differentialprojectionof two of the inferiorparietal lobuleof macaque cytoarchitectonic subregions uponthe deeplayersof - 254 colliculus . Journal . , 235 , 241 of Comparative superior Neurology Maruff : Currie , P., & , J. (1995 with Alzheimer ' s disease . ). An attentional graspreflexin patients - 701 . , 33 (6), 689 Neuropsychologia Mesulam networkfor directed , M. M. (1981 attention andunilateral . Annals ). A cortical neglect - 325 . , 10 , 309 of Neurology : Desimone attention visualprocessing Moran , J., & in the extrastriate , R. (1985 ). Selective gates - 784 cortex . Science . , 229 , 782 Morrison : Foote . L. (1986 and serotonergic , J. H., & innervationof cortical ,S ). Adrenergic , thalamical andtectalvisualstructures in Old andNew World monkeys . Journal of Comparative - 138 . , 243 , 117 Neurology

Introduction to Neural Networks

. , Beau : Evans Murtha . A., Chertkow , R., Dixon ,R , M., & , A. (1996 ). Anticipationcauses regard - 112 increased bloodRowto theanterior . Human BrainMapping . , 4, 103 cingulate Niitinen , R. (1992 andbrain . Hinsdale : Erlbaum . ). Attention , NJ function andposterior attention in hemispheric networks , A. (1991 ). Roleof anterior Nakagawa asymmetries - 321 decisions . Journal Neuroscience . , 3(4), 313 of Cognitive duringlexical NormanD. A., & : Shallice . Winedandautomatic controlof behavior , T. (1986 ). Attentionto action -regulation . In R . J. Davidson . Schwartz : D. Shapiro .), Consciousness andself , G. E ,& (Eds : Plenum Press . (pp. 1- 17 ). New York Pardo : Raichle cortex , J. V., Pardo , P. J., Janer , K. W., & , M. E. (1990 ). The anteriorcingulate mediates selection in the Stroop attentionalconOidparadigm . Proceedings of the processing National - 259 . , 87 , 256 Acadmry of Sciences . E., Fox : Raichle emission Petersen ,S , P. T., Posner , M. I., Mintun , M., & , M. E. (1988 ). Positron - 588 studies of the cortical of -word processing . Nature . , 331 , 585 anatomy single tomography : Raichle emission Petersen , S. E., Fox , P. T., Posner , M. I., Mintun , M. A., & , M. E. (1989 ). Positron studies of the processing of singlewords . Journal Neuroscience ,1 , of Cognitive tomography - 170 153 . Posner . Hillsdale : Erlbaum . . M. I. (1978 . NJ ). Chronometric of mind Erplorations Posner andfunctions of selective attention . In T. Boll & .). : B. Bryant(Eds . M. I. (1988 ). Structure - 202 MasterLecture in Clinical andBrainFunctions . DC: (pp. 171 ). Washington Neuropsychology American Association . Psychological - 79 Posner : Dehaene networks . Trends in Neuroscience . . M. I.. & . S. (1994 . 17 ). Attentional (2). 75 Posner : Petersen of the human brain . AnnualReview . M. I.. & . S. E. (1990 ). Theattentionsystem - 42. . 13 . 25 of Neuroscience . E.. Fox Posner : Raichle . (1988 of cognitivefunctions ,S . M. E . M. I.. Petersen . P. T.. & ). Localization - 1631 in thehuman . Science . brain . 240 . 1627 Posner : Presti attentionand cognitivecontrol . Trends in Neuroscience . M. I.. & . D. (1987 ). Selective - 17 . . 10 . 12 Posner : Raichle . NewYork: Scientific American . . M. I.. & . M. E. (1994 ). Images of mind Library Posner : Shulman automatic ? . M. I.. Sandson . J.. Dhawan . M.. & . G. L. (1989 ). Is word recognition A cognitive anatomical . Neuroscience 1 5060. . . Journal of Cognitive approach Raichle . A . Videen : Petersen . M. E.. Fiez .J . T. 0 .. Macleod . A.-M. K.. Pardo . J. V.. Fox . P. T.. & . -related in human S. E. (1994 brainfunctional learning ). Practice changes anatomy duringnonrnotor . Cerebral Cortex . 4. 8- 26. -parietal .. & Robert son : Knight of lesions of temporal . L. Lamb . M. R . R. T. (1988 ). Effects junction - 3769 on perceptual andattentional in humans . Neuroscience 8 3757 . . Journal of processing . (1984 Roland of motorcontrolby the normalhuman . Human brain Neuro . P. E ). Organization - 216 . . 2. 205 biology : Coles .. M. G. H. (Eds .) (1995 . Oxford : Oxford University ). Electrophysiology of mind Rugs . M. D.. & Press . : Berg source Handbook . Version2.0. Neuro Scan ). Brainelectrical . analysis Scherg , M.. & , P. (1995 - 6103 .. 10Pidgeon Inc Hill Drive . . Suite100 . Sterling . VA 20165 Shallice : Burgess . T.. & . P. W. (1991a ). Deficitsin strategyapplications following frontal lobe in man . Brain 114 727 741 . . . damage

Posner & 1 : Badgaiyan : AttentionandNeuralNetworks

Shallice , T., 6: Burgess , P. (1991b ). Higher-order cognitive impainnents and &ontallobe lesions in man. In H. S. Levin .), FronfQllobe , H. M . Eisenberg , 6: A L. Benton (Eds functionand dysfunction . (pp. 125- 138). New York: Oxford University Press , G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Simpson Bulletin , 96, 316- 340. , A ., Abdullaev, Y. G., Posner : Raichle , M . I., 6 , M . E. (1995). Scalp electrical potentials Snyder reflect regional cerebralblood flow responses during processingof written words. Proceedings of the Nation Jll Acadmryof Sciences , 92, 1689- 1693. . Journal of Erpm Stroop, J. R. (1935). Studies of interference in serial verbal reactions ' mental , 18 , 643- 662. Psychology Swanson , J. M ., Posner , M . I., Potkin , S., Bonforte, S., Youpa , D., Fiore, C., Cantwell, D., 6: Crinella , F. (1991). Activating tasksfor the study of visual-spatial attention in ADHO children: A . Journalof Child Neurology , 6( Suppl.), sl19 - 127. cognitive anatomicapproach TuckerD . M ., Liotti , M ., Potts, G. F., Russell , G. S., 6: Posner , M . I. (1994). Spatiotemporalanalysis of brain electricalfields. HumanBrain Mapping, 1(2), 134- 152. Van Essen : Felleman , D. C., Anderson , C. H., 6 , D. J. (1992). Information processingin the primate visual system : An integrated systemsperspective . Science , 255, 419- 423. Van Essen : Olshausen , D. C., Anderson, C. H., 6 , B. A . (1994). Dynamic routing strategiesin -scaleneuron motor, and cognitive processing . In C. Koch 6: J. L. Davis ( Eds , .), Large Jll sensory theories . , MA : MIT Press of the brain (pp. 271- 300). Cambridge ' : Zappala Villardita, C., Smirni, P., 6 . Archivesof , G. (1983). Visual neglect in Parkinsons disease , 40, 737 739. Neurology Whitehead , R. (1991). Right hemisphere . processingsuperiority during sustainedvisual attention , 3(4), 329- 334. Journalof CognitiveNeuroscience Wilkins, A . J., Sha Uice . (1987 , T., 6: McCarthy, R ) . Frontal lesionsand sustainedattention. Neuro25, 359 366. , psychologia

Introduction to Neural Networks

ANeural Network Model ofMemory , Amnesia Cortico ,and Hippocampal Interactions

Combined neuropsychological and biological evidence suggests a fundamental but selective role for the hippocampal system (Hs ) in some forms of learning . Hs is necessary for rapidly (eventually , in one exposure ) forming declarative , explicit long - term memories , whereas it appears to be unnecessary for the progressive acquisition of procedural , implicit memories . Recall of previously acquired declarative memories gradually becomes independent of Hs itself . This suggests a graded process whereby traces are consolidated and stored in another structure , such as the cerebral or cerebellar cortex . In the study of declarative explicit memory , and particularly of the role of Hs in this type of memory , human neuropsychology is historically ahead of animal neurophysiology . This unusual state of affairs raises the problem of transposing to the animal domain concepts specifically coined for human cognitive functions . Integrating in a single model concepts from neuropsychology and animal neurophysiology implicitly assumes continuity among memory processes occurring both in human and nonhuman primates or other species.

Anterogradeand RetrogradeAmnesia
Most authors agree that a major milestone in memory researchwas a result published by Scoville and Milner in 1957. They observed a dramatic but selectivememory impairment following a bilateral ablation of hippocampus and related structures in the medial temporal lobe to control an otherwise intractable epilepsy. The patient HiM .' s severe memory loss can be contrasted with preserved skill learning and priming effects , including broad labeled as Schacter , implicit memory ( , Chu, & Ochsner , capacities procedural 1993). Most authors draw two conclusionsfrom theseresults : 1. Some (declarative ) forms of memory are initially stored in Hs, and then transferred to other, more permanent storage sites , such as the gradually neocortex. 2. Other (procedural . ) forms of memory are directly stored in cortical areas

. But These statementscapture the gist of many neuropsychological results in the results . Recent results neuro do not , fully explain mostly they in neural and enable more but also modeling , imaging precise physiology, statementsabout the locus and nature of stored information. It should be noted, however, that the precisemodalities of memory consolidation remain a mystery.

Buff~r
' : O Keefe Unlike some neurophysiologists (Burgess , 1994; Horel, , Recce , & ' that Hs does not we : Nadel, 1978) and modelers 1994; O Keefe & , suggest of specific playa central role in the primitive storage and recall of the content or link operators among compressed episodesand events. Only the connectivities of cortical activation patterns are transiently hippocampalrepresentations stored in Hs. These patterns of connectivity are transferred to and developed in the cortex under Hs control. Full-fledged memory traces are initiated and finally stored at the cortical level. There is no capacity for transfer and storage of full traces in Hs. Preliminary evidence suggeststhat procedural memory similarly involves complex cortico-subcortical circuits but does not rely on the Hs for consolidation. The basalganglia, in particular , appear to be involved in these cortico-subcortical circuits whenever es. are implicated in learnedprocess motor responses Whatever the cortico- hippocampal mechanismsinvolved in long-term consolidation of declarativememories , experimental results suggest that the role of the Hs extends far beyond that of a transient long-term memory (LTM ) store, during the time required for cortical trace consolidation. It seemsto be implicated during the very first wave of cortical processingtriggered by stimulus input . Neuropsychologicalevidencefrom amnesicpatients suggeststhat even in the simple role of holding and buffering information, Hs is necessaryin the shorl-tenn range , as far as information exceedsthe short-term memory (STM) . This should not be surprising sincethe Hs activation lag in responseto span an external input does not exceeda few hundred millisecondsrelative to primary . A strong conclusion is that both systemsprocessin parallel cortical areas and interactively. Similar evidencearguesin favor of a crucial Hs role in the intennediate ) buffer. rangeas a Working Memory ( WM

' The Hs s multi range temporal capacity makes it a buffer that can reenact activation patterns of information at cortical levels. This occurs during the lengthy process of LTM consolidation and also during the moment-by . It functions as an intermediate register moment operation of current tasks order of minutes the on ) supporting WM . This last hypothetical (lasting

Introduction to Neural Networks

function constitutes an automatic working memory. The two proposed functions are complementary ; the process of LTM consolidation is engaged only when WM processing has transformed a short -lived trace to a transient LTM trace. Also , they are based on similar physiological processes, namely reenactment of cortical , electrical patterns either spontaneously or in reaction to a cortical cue. This unique process allows network activation , according to its locus of initiation , to trigger buffering and rapid learning of information in hippocampal subsystems (if the initial focus of activation is cortical ) or to reactivate cortical patterns , reinstating recent memories (if the initial focus of activation is endogenous in the hippo campus). The reactivation of cortical activity patterns derives from the capacity of the Hs to function according to two distinct , complementary modes: the read mode registers and processes external , cortical information ; the print mode " " or reactively reinstates the corresponding patterns of endogenously activation locally in the hippocampal subsystems, or in both the Hs and the cortex . These two modes correspond to two clearly defined electro physiological patterns in some animal species: theta and sharp waves (Buzsaki, 1989 ). In primates , only sharp waves have been consistently identified . This peculiarity may correspond : first , to a spatial localization of theta activity , which would not affect the entire hippocampal network , and second, to a functional extension of the read mode which would not be limited to the physical exploration of the enviornment but participate as a follow -up to any " " endogenously generated cognitive processing . In our model , the reactivation " function is related to the internal bursting " capacity of the CA3 pyramidal neurons , and their collective capacity to synchronize under the modulation of septal inputs .

Consistent with its transient buffer function, behavioral and neurophysiological results suggestthat the Hs plays a cognitive role in learning the temporal order of serial sequences . This , at least at a low level of processing function may be involved in the processof map formation in lower species . In addition, the conditioning literature suggeststhat some parts of the hippocampus play an important role in learning temporal intervals between events . Some of these functions may be based on very basic process es of differential synaptic plasticity (Granger , Whitson, Larson , et al., 1994) and on the correlational learning capacitiesattributed to the hippocampal subsystems events (Banquet , as they are applied to successive , Gaussier , Moga , et al., 1997a ). Some other functions, such as timing , may result from population . coding by cells endowed with different dynamics In this chapter we review evidence for the term , , declarativememory long consolidation function attributed to long-term potentiation (LTP). Inaddition , we argue that the automatic WM function is supported by an intermediate transient memory (ITM ) operating in parallel with the cortical WM .

, and Cortico-HippocampalInteractions Banquetet al.: Memory , Amnesia

We compare experimental results in both human and nonhuman primates . We contrast the WM function with the more classic LTM consolidation function attributed to Hs. We also provide physiological and behavioral evidence for Hs involvement in temporal order sequence processingand timing . This forms the basisof our neural network model implementation .

HIPPOCAMPUS IN LEARNINGAND LONG-TERM MEMORY CONSOLIDADON


We propose that the Hs' s memory function covers the entire temporal spectrum " LTM from STM to transient LTM , but does not include " permanent . One characteristicof Hs is the transient characterof storage there. Arguments in favor of the classicLTM consolidation and WM functions will be . presentedsuccessively We present a brief review of neuropsychology and neurophysiology of amnesiafollowing medial temporal lobe lesion. This literature suggeststhat a transient LTM function of the Hs plays a causalrole in LTM consolidation and learning. This role will be contrasted with a cortical function that involves mostly STM and permanentLTM registers . . We use the term hippo campus proper(H ) to refer specificallyto the CA field ' Ammon s horn . ( ) . Hippocampal formation (Hf ) includes the CA field, dentate gyrus (Og), and subiculum . . Hippocampal system(Hs) further includesthe parahippocampal region comprising entorhinal cortex (Ec ) as part of the parahippocampalgyrus, perirhinal and parahippocampal cortices . From the vast body of experimental literature on Hs' s role in memory and , learning, we report only those results that are most relevant to our purpose those from . a For more detailed review of the particularly neuropsychology , Otto , and Cohen (1994). experimentalanimal literature, seeEichenbaum Our focus is the finding that a patient with an Hs lesion is unable to learn certain new information (anterograde amnesia ) and displays graded forgetting for the most recent past precedingthe lesion (retrograde amnesia ).
Anterograde Amnesia

Hs lesions produce a profound de6cit in learning new material but leave intact performanceand recall based on material acquired some time before the lesion (the delay varying according to speciesand subjects ). Scoville and Milner (1957) and Milner (1966) provided the 6rst evidencefor anterograde amnesia . They describedthe patient HiM ., following removal of the medial temporal lobe (Hs) for the control of an otherwise intractable epilepsy. The authors proposed a relationship between the extent of the damageand the . severity of the consequentamnesia

Introdudion to Neural Networks

More recent results &om selective lesions in nonhuman primates and also postmortem anatomicalanalyses&om humanshave confirmed that the magnitude of the memory deficit is related to the extent of the lesion. Further , anatomically selective lesions have shown that the very nature of the functional deficit (temporal or correlational ) changeswith the subsystem(para et al., 1994). hippocampalregion or Hf ) affectedby the lesion (Eichenbaum
Declarative and Procedural Memory

Whatever the extent and nature of Hs lesions , only some forms of new learning are affected. Classically , Hs is believed to be necessary for the ability to form rapid new associations for elements that one can "bring to mind ," recognize as memory events , and explic.itly describe. These memories can be flexibly used (in new contexts ) to govern subsequent behavior . However , we argue that this close association between declarative memory and Hs function does only partial justice to the results. In particular , we argue that the declarative quality of this type of memory depends more on the cortex , the final storage site of the memory , than it does on the hippo campus itself . On the basis of experimental results , it seems likely that the Hs is involved in correlational learning, that is, in the formation of new associations between multiple events co -occurring in a spatiotemporal scene, where the temporal dimension is a crucial parameter taken into account by the Hs. The temporal dimension of to - be-memorized information seems even more critical than the spatial dimension for the mandatory involvement of " " Hs. Here temporal dimension means that correlational information cannot be captured in a single snapshot , as in a simple spatial scene. Either several consecutive views of a complex scene must be taken , or the total information must be delivered in sequential parts , as in a sentence delivered word by word on a screen, or in the unfolding of a melody . This may be the reason why trace conditioning is affected by Hs lesions. Trace conditioning does not involve multidimensional stimuli , but it does involve a delay between the delivery of conditioned (CS) and unconditioned stimuli (US). Conversely , delay conditioning is not affected by Hs lesions. Delay conditioning involves a temporal dimension , but the CS and US overlap . The declarative memory system refers specifically to LTM , and includes " " episodic memories as defined by Tulving ( 1983 ), that is, memories for the specific content of events that occur only once. Declarative memory also includes other subsystems in Tulving ' s taxonomy , such as semantic memory . In semantic memory , the factual memory aspect appears to be affected by Hs lesions, but word associations and concept or rule formation are less affected. For example , artificial grammars can be learned and used after Hs lesions. " This suggests that the declarativeness " of the memory is not as critical as whether or not information can be acquired by practice . Whenever practice is possible in the learning process, Hs becomes unnecessary. Episodic

, and Cortico-HippocampalInteractions Banquetet al.: Memory, Amnesia

memory is just the most typical kind of declarativememory and it cannot be . Conversely learned by practice , procedural memories are characterizedby the ability to learn by practice , relative inaccessibility to conscious recollection . These featuresare or declarativerecognition, and inflexibility of use the opposite of those that characterizedeclarative memory. Notably , they are almost completely unaffectedby Hs lesions (Schacter& Graf 1986). This . suggeststhat much practical learning is sparedin hippocampalamnesia

RetrogradeAmnesia
One characteristicof patients with Hs lesions is a temporally graded and limited retrograde amnesia . This contrasts with an ungraded and extensive amnesia following more global brain injuries involving extensive cortical . Retrograde amnesiacan be defined as memory loss for events preceding damage . It was discovered only secondarily in hippothe onset of amnesia campal patients. Some authors question the existence of this deficit : Weiskrantz : McCarthy, 1988; Warrington & , 1978). (Warrington & in The first indication of retrograde amnesia Hs lesions came from the patient HiM . who had a selective deficit for material acquired two or three : Milner , 1957). Although old material is years prior to surgery (Scoville & . A forgetting a selective deficit for prelesion memories there is , generally spared two the can be established between the , although specific gradient are highly controversial. tests used for this assessment The severity of retrograde amnesiais highly variable both between species and from one patient to the next. Its severity depends , as is the casein In of the lesion. amnesia on the extent , patients with Hs damage anterograde , retrograde amnesia spans one or two years prior to surgery (Zola: Amaral, 1986). With more extensive lesions , retrograde ,& Morgan, Squire : Squire amnesiamay span at least two decades(Mackinnon & , 1989; Squire , HiM an this rule considering . Thus . is to Haist, & : Shimamura 1989 , , ) exception his relatively limited retrograde amnesiadespite an extensive lesion. , performed at different periods Prospectiveexperimentson rats and monkeys before the critical Hs lesion, have confirmed the retrograde memory deficit : Fanselow : Jaffard . Different authors (Cho, Beracochea , & , 1992; Kim & , have severe : Squire 1992; Winocur, 1990; Zola Morgan & , 1990) , reported reliable impairments for memories acquired up to four weeks prior to surgery in monkeys , and up to a few days in rats. amnesia tells us that Hs is necessaryin order to memorize Anterograde certain new information. Graded retrograde amnesiasuggestshow Hs interferes . Together, these results provide with the LTM consolidation process evidence for a gradual process whereby Hs is necessaryto organize and consolidate memories at the cortical (or cerebellar ) level. These memories of the medial eventually becomeindependent temporal lobe. This does not exclude the possibility that Hs is involved in their reactivation or retrieval.

Introduction to Neural Networks

INTERMEDIATE MEMORY SUPPORTING AN AUTOMADC WORKING MEMORY COMPONENT


The existenceof a graded retrograde amnesiais a strong argument in favor of a transient LTM probably basedon hippocampalLTP. We now argue that WM is basedon both a cortical systems) and a hippocampalITM . We present neuropsychologicaland brain imagery arguments in favor of STM and ITM functions of the hippocampus .

PsychologicalArguments for an Automatic Working Memory


Component

Baddeley(1986) proposed a multicomponent WM model as a substitute for ' Atkinson and Shiffrin s (1968) modal model. The modal model implied a unique STM store as a necessarypassageto LTM . However, a single STM store could not simultaneously function as an adequateWM ; therefore , it evolved into a multicomponent model (figure 4.1). Still, Baddeley viewed WM as a single common resource . , with a limited capacity

Sensory registers Short t erm store Visual S TS ( ) I I t erm stor Long Input L TS ( ) I Auditory I Tempora IU ~ _ . I -I working memory Havtic

Figure 4.1 The Row of information through the diHerent memory systems in the model of Atkinson and Shiffrin (1968) and the working memory of Baddeley(1986).

et aI .: Memory - Hippocampai Interadions , Amnesia , andCortico Banquet

Working Memory as a Cortical System


" The definition of Working Memory as temporary storage of information " is in connectionwith performing other, more complex tasks generalenough to allow for many extensions and modifications of the model. Baddeley assumesa limited-capacity attentional controller, the central executive , that two slave . The modality specificsystems articulatorylooprehearses supervises -basedinformation and manipulatesmemory for sounds . It comprises speech a memory store for holding phonological information for a period of 1 to 2 seconds , coupled with an articulatory control process (Baddeley , 1986). Overt or covert subvocal articulation refresh es auditory memory tracesand feedsphonologically translatedvisual information to the phonological store. The visuospatial holds and manipulatesvisual and spatial images . sketchpad The temporary storage of visual information in a visuospatial sketchpad , a parietal system implies an occipital system involved in visual aspects involved in spatial coding, and possibly some frontal lobe participation , 1988). A related construct was proposed by Fuster (1995), (Goldman-Rakic called active memory . This is a state rather than a memory system and includes a widely distributed and changing representationalnetwork in the awake organism . Active memory includes WM , but does not presuppose any mental or cognitive operation. WM function appearsto be under consciouscontrol; therefore , it is likely to have a cortical location. The same is also assumed for animal WM , 1988). However, the underlying brain structures that support (Goldman-Rakic the executive and slave systemsare not fully understood . More recent ' versions of Baddeleys theory have incorporated results from brain imagery and animal neurophysiology (Baddeley , 1995). Unfortunately, most of the to date metabolic or electrical brain imaging with , involving experiments event related potentials (ERPs ), use paradigms that were designed to study ' componentsof Baddeleys (1986) model, the rehearsalsystemsin particular : Frackowiak , Frith, & , 1993). Moreover, the ERPapproach(Ruchkin , (Paulesu et al. al. Canoune 1991 Ruchkin Canoune et is 1992 , , , ; , Johnson , , , Johnson ) better suited for exploring the cortical mantle than the deep structures . Therefore , thesestudiesconfirm the involvement of cortical structuresin different WM paradigms . Yet, classicrecall paradigms , not formally requiring rehearsal , show joint cortical and hippocampalactivation (Squire , Ojemann, Miezin, et al., 1992).
Working Memory as a Hippocampal System

Neuropsychologicaland imaging results suggest both an automatic component of WM supported by an intermediateregister located in Hs and a controlled cortical component of the WM system , explored both in humans : Alexander, 1971). Like the cortically , 1986) and in animals (Fuster & (Baddeley controlled WM , this system has both storing and processingcapacity .

Introduction

to Neural Networks

It is based on a complementary set of intrahippocampal and hippocampo cortical loops. These are combined with a battery of memory registers covering a large temporal spectrumand basedon transient synaptic facilitation demonstrated in different hippocampal subsystems (Buzsaki , 1988; Jones , 1993). Moreover, the operation mode of the intermediatesystem supporting WM is restricted to the intermediate range (minutes ) of the more comprehensive LTM consolidation processgenerally attributed to Hs. This general process consists of maintenanceor reenactmentof cortical activation patterns . A consequence by reverberantactivity between reciprocally connectedsystems of this hypothesis is that the existing dichotomy in LTM between declarative explicit memory and procedural implicit memory is extended to WM . Psychologicalargumentsin favor of the hippocampalautomatic component of WM are provided below. Refreshing either the visuospatial sketchpad or the phonological store . This is also true of by rehearsalevery secondor so is a controlled process task verbal that is the concurrent , cognitive , usually , required from subjects . Supposedly in WM paradigms , the central executive responsiblefor planning , and coordination of information monitors both of , strategy selection es work in parallel. Shiffrin and these tasks . Thus, several controlled process Schneider(1977) demonstratedthe very limited capacity for controlled processing in the human brain. Only severalautomatic process es , or at best, one es controlled and severalautomatic process , can be performed in parallel. Yet, WM are tractable without overwhelming used to test the , surprisingly paradigms or even , by patients difficulty aging subjects (Brebion, 1994). There may be several reasonsfor this, not necessarilymutually exclusive , . For example and more or less implicated according to tasks and subjects , WM may be largely automatizedin a complex cognitive task used to assess spite of its complexity, especially when it involves verbal comprehension . Thus, attention is relatively free to focus on active rehearsalof to-bememorized material. Alternatively , to-be-rememberedinformation may be more or less related to the cognitive task. Therefore , competition between not be involved functions and ; rather, it may be may holding processing tasks . This of the two to type paradigm (where to-beremembered integrate possible information is related to the cognitive process in progress ) is close to an actual WM involvement in natural conditions. . But we Certainly one cannot deny the reality of the rehearsalprocess an in terms of limited control led is that it , expensive ( processing suggest in most everyday ), low -level, and rote strategy that is not necessary capacity situations. Whenever possible , subjects use cognitive strategies to create items to increase the limited capacity of STM. chunks of supraordinate . We therefore propose the existence However, this is still a controlled process WM an ITM to . of an automatic support , , that does not need rehearsal ITM is in a functional relationship with the various cortical areassuch that it es recent memoriesrelevant to the task in progressby simple readily refresh maintenance or reenactment of the corresponding activation patterns in

, and Cortico-HippocampaiInteractions Banquetet al.: Memory, Amnesia

cortical populations. This reactivation does not preclude a state of priming or subliminal activation of the cortical areas . This function is exactly what Hs in the different es of supposedlyperforms long term consolidation process declarative episodic and factual semantic memory traces . Both WM and transient LTM involve episodic , context-dependent , or factual types of tobe-memorizedinformation. The only important distinction between the two is that one trace will have been " long-term- potentiated" and transferred to LTM , whereasthe other (and eventually the activity pattern) in WM is only a candidate for LTP and must undergo a test of eligibility for permanent LTM storage . We further hypothesize that the two proesses , LTM consolidation on the one.hand, and WM refreshing on the other, are roughly similar and complementary . Only the duration during which the two types of memory traces are relevant for the brain differs. This is on the order of minutes for W Ms , " and weeksor months for memoriesthat must be " permanently consolidated in the long-term stores . They are complementarybecause transformationinto transient LTM can only be consideredwhen the long-term relevanceof the to-be-stored information has been confirmed by processing in the intermediate , or in both. The hippocampal store, in the cortical WM system of neurophysiological counterpart information selection for long-term storage correspondsto the transition &om loop-iterative and punctual activation or &om short-term synaptic facilitation (based on short-term potentiation, STP ) to a transient long-term synaptic facilitation (basedon LTP). Both types of learning are documented at different levels of hippocampal subsystems , 1988; Jones (Buzsaki , 1993).
Neuropsychological Arguments

Neuropsychologicaldata remain important, sincebrain imaging experiments . First, Brown (1958) and performed during WM paradigms are still scarce Peterson and Peterson(1959) showed that holding information in STM is and that information is lost rapidly if active rehearsal dependenton rehearsal is prevented. This fact distinguishes primary memory (James , 1890) resulting &om a STM extension due to the rehearsalprocess&om immediate STM. ' Second , normal subjects responsesto different recall and recognition tests show, in the absenceof rehearsal , a residual memory, that is, the amount to which STM decays asymptotically. This residual memory is suppressed after a bilateral Hs lesion. Therefore , we attribute it to a hippocampal ITM . component , in amnesiafollowing a medial temporal lobe lesion, STM is Classically in intact , contrast to the complete loss of declarativeLTM (Baddeley& fully , 1992). The picture emerging &om the Warrington, 1970; Cave & Squire study of amnesicpatients appearsmore subtle than this, particularly when one considers how the extent of the lesions influences the depth of the deficit.

Introduction to Neural Networks

The following neuropsychologicalarguments derive &om recall and forgetting curves and visual recognition obtained by classictests of STM and LTM , in both normals and amnesicpatients. Patientsshow a deficit in learning and recall for verbal as well as nonverbal material. In terms of our model, these tests explore both STM and ITM , according to the variable delay of recall. Sincethe depth and nature of a memory deficit depend on the extent and location of an Hs lesion, these parametersare taken into account in the . Bilateralextensive lesions with complete loss of interpretation of the results function will be treated separately &om unilateral lesions and hippocampal partial loss of hippocampal function. But the most important parameter for separatingthe hippocampal&om the cortical componentsof WM is whether or not subjectsengagein controlled rehearsal , such as that performed by the 's WM in model. 1986 ) articulatory loop Baddeley (
Allowed Rehearsal

The results from subjects with complete loss of Hs function (like patient HiM .) will be emphasizedsince normal, or close to normal, performanceon tests of STM or ITM implies normal performancewhen hippocampal function is partially preserved . In this case , we attribute the corresponding performance to STM or to the controlled, cortical component of WM , since rehearsalis allowed and Hs function is lost. This complete loss of Hs function results from bilateral and extensive resection of the medial temporal ' lobe, as in HiM . s case , or from unilateral resection associatedwith severe ) or with a severe dysfunction of the degeneration (postmortem diagnosis P.B. and H.F.). medial hemilobe (respectively cases spared temporal In the first set of paradigms , rehearsalis natural, involving verbal material (consonant trigrams) well within the memory span and presentedin a variable , Stoddard , 8t Mohr , 1968). STM delay matching-to -sampletask (Sidman in William can be extended at will what James(1890) called a primary range memory (figure 4.20 ), even for patients with total loss of hippocampal function. In the second set of paradigms , to -be-rememberedmaterial is easily verbalized but is at the limit of the memory span , as in the short version of the of visual maze(Milner et al., 1968). This test requiresmemory for a sequence turns (eight choice points). Patients cannot learn the task even after many trials. Thus, the patients have a cognitive deficit related to their inability to ) simultaneously maintain and organize information (type of turns to make and implement the task (topographic translation of the turns on the maze ). This is a genuine WM task in which rehearsalis not sufficient becausethe . Furthermore , practice does not memory load is at the limit of the STM span . over 125 trials, as if There is no evidence of learning improve performance the subject were unable to devise a learning strategy in order to split a difficult . task into accessible subgoals

- Hippocarnpal Interactions et al.: Memory . andCortico . Amnesia Banquet

100 _ SICS A --0 90 -- CWfIa .S 60 70 lV 60 ~ ", ----d ! f v .Q , 50 . . . \', o---a IZ ,: '" ' 40 "' : ~ -_ b _ _ '" ' 30 Q "," .': a . d 20 1 := :4~5:::::: ] 1 2:~3 . 10 . :0 6 7 6 9 ITEM NUMBERS

100 90 80 70 60 50 40 30 20 10 0DE 30 6 ) ec s (
- - 0- _ SICS C , . , RO .S

100 - _ SICS C -- CWfIa .S 90 --0 60 0 \" 70 """ "" 60 "" \Q 50 ."" '"' 40 "'" 30 "~----------o 20 10 0 5 6 7 LEN6TH ' STRIN6

-recalltask - component free . Mean percentage correctrecall(immediate Figure 4.2 (A) Two recall for subjects ) asa functionof orderof presentation lesions of the hippo presenting partial 's -tenn forgetting or Korsakoff -tenn . (8) Sort (amnesic ) and controls . Short syndrome campus retention of word triadsfor amnesic andnonnalcontrols . (C) Digit span : Immediate for memory andnonnalcontrols . Meanpercentage of sequences digitsin amnesic correct asa function patients of sequence . ( 0) Delayed . (Adaptedfrom Baddeley and length matdtingadjusting delay .) , 1910 , with pennission Warrington

In the third set of paradigms , the to-be-memorized material is not easily verbalized (elliptic geometricforms with one variable radius to compareto a ). In this case , extension of the STM range is not sampleafter various delays for see 4.20 . Even with very sensitive measures ( ) possible patients figure ,a limited residual control of the sample stimulus on performanceis restricted to the 16 to 24-second range , the classicdecrementalSTM range. Remark

Introduction to Neural Networks

fL ~
TAn~

S9NI

D El A YE DMAT C Hlta A D Jtmlta DElAY

B ' H . M . J1 ~ S 4 " \ J ' 3 Jru " I Ja J . 1 0 ~ )" \ :~ : ~ SV TRw .

ably, this is not the casefor normal subjectsor even for nine to twelve yearold children. Both groups show no performancedeterioration at delays up to . Poor performancein the caseof extensive bilateral lesions has 40 seconds also been found in tests involving other types of geometric items (Milner et al., 1968). Finally, evidencefrom classicdigit span tests further extend these deficits of various lengths are presented at a to immediate STM. Digit sequences . second rate of one Subjectsnamethe recalleddigits at their digit per typical . The percentageof strings correctly recalled (figure 4.2C) has an own pace inverse linear relation to the length of the string, and is similar in amnesics exceeding and controls. But amnesics consistently perform worse on sequences : Warthe normal memory span(approximately sevenitems) (Baddeley& to another : Arbit , 1966). These results point rington, 1970; Drachman & critical factor in recall performance besides the delay of retention , namely, an lobe seems to medial . The load important role with memory play temporal Furthermore load. and duration to both , this result memory delay respect of information the overflow shows that the medial temporal lobe helps to buffer out of the very limited capacity , STM store. This suggests , attended that rehearsalis no longer efficient as a consequenceof the overflow of . Still, normal subjects demonstrate residual the immediate memory span in mnemonicperformance spite of this overflow. Taken together, these results confirm the integrity of STM for material . They also confirm the efficiency within the immediateSTM spanin amnesics of the articulatory loop as a rehearsaldevice for extending the immediate . This STM span into a primary memory for material that is easily verbalized , the articulatory loop is intact and efficient in Hs patients. Nevertheless results also suggestthat rehearsalis not enough to guaranteerecall or learning in patients . If the material is difficult to verbalize or exceedsthe immediate , as in the short version of the visual maze (Milner et al., memory span 1968), the memory processis completely disrupted. Therefore , patients with subgoals Hs lesions find it difficult to break down a complex task into accessible . Conversely , the results do not provide strong support for a corresponding rehearsalsystem involved in refreshing the visual sketchpadwhen . If that were the case material is not easily verbalized , then patients would show normal performancein geometric tasks , which they do not.
Rehearsal Prevented Tests in which rehearsal is prevented during the delay between learning and recall resemble experiments designed by Baddeley to test WM . Indeed , they combine holding information with an interfering task. However , they differ from genuine WM paradigms in that the to -be- remembered items are not related to the interfering task, nor are they related among themselves . Conversely , in a WM paradigm , the task is not designed to interfere with

, and Carlita- HippocampalInteractions Banquetet al.: Memory . Amnesia

memorizing the items. The items may even be related to the task, and task . . Also, in a typical WM task performanceis quantitatively evaluated , STM and ITM componentscannot be dissociated . These componentscan be dissociated in memory tests by monotonic variation of the recall delay. The results of thesetests dependcritically on the extent of the lesion. Bilateral Lesions of Hs Subjects with bilateral extensive lesions of the medial temporal lobe, like the patient HiM ., are in a severe , even dramatic , condition. They depend on continuous rehearsalof to -be-rememberedinformation . Catastrophic forgetting is induced by distraction (Milner , 1966; Milner et al., 1968). The most dramatic illustration of this psychological condition was reported by Milner (1959): H.M . was able, by devising an elaborate mnemonic scheme , to remember a three-digit number for fifteen minutes but it as soon as he was distracted , . Why then does a normal forgot subject have no difficulty remembering such a number after maintaining it for fifteen minutes , even after being distracted ? The hypothesis that fifteen " L minutes is enough exposurefor a " permanent TM storage is not tenable . There is little chancethat transient storage takes place at the cortical level; otherwise the patient HiM . should have learned this information, as he learnedinformation in proceduralmemory paradigms . Sincethe main differencebetween normals and patients in this caseis the of the Hs, it is logical to attribute the primary causeof presenceor absence the recall deficit to the Hs and not to the cortex. This provides no information about the actual mechanisms responsiblefor the storage and retrieval of information in normals . In normal subjects , continuous rehearsalcannot be for effortless recall since the very definition of distraction directly responsible , . Recall implies an interruption of the controlled rehearsalprocess in normals (compared to nonretrieval in patients with extensive Hs lesions ) strongly suggests that continuous rehearsal somehow succeedsin rapidly , not at the cortical level, but at the Hs level. Our ITM laying down a trace hypothesis explains why normal subjects retain this information even is interrupted and the information is irrelevant to long-term though rehearsal storage (LTM consolidation was not yet involved). Fifteen minutes is well within the WM range . This is, therefore , an argument in favor of an automatic of WM . We will seehow recent findings on the multiplecomponent range memory registers operating in the hippocampal subsystems , in conjunction -loop system of the with the complex closed , may hippocampus provide information about the actual mechanismsresponsible for these -range remembrances transient intermediate (which are very useful for our moment to moment processingof the continuous information flow ). A more formal assessment of this type of patient was performed in a visual task based on face memory recognition. Subjects were shown twelve facesand 90 secondslater had to selectthesefacesfrom an array of twenty ' five. Patients performancefell to chancelevel when a distracting task was

Introduction to Neural Networks

interpolated between presentation of the two sets of photographs (Milner et al., 1968 ). We have seen previously how an excess in memory load , or difficult verbalization of to -be- remembered material , is equivalent to preventing rehearsal.

Partial Lesion of Hs The picture is apparently different when there is of hippocampalfunction. This is the casein patients after partial preservation unilateral temporallobectomy , or in alcoholic patients suffering horn Korsak ' . Both groups show an identical pattern of results (Baddeley& off s syndrome : , delayed recall tests with interfering Warrington, 1970). As we have seen tasks are similar in some respects to WM paradigms . The variable delay between item presentation and recall allows a quantitative separation between deficits in STM (lessthan 20 to 30 seconds ), and those in intermediate ). This is not the casein WM paradigms memory (beyond this range , which involve a more natural situation, in the sense that the to-be-memorized material may be related to the associated processingtask. -term forgetting task, subjects receive item In the short sequences(three words) well within the memory span . These are presented for 3 seconds . after delays varying according Subjectsare askedto recall the item sequences to trials (0, 5, 10, 15, 30, or 60 seconds ). During thesedelays , they perform a control led . The forgetting tightly intervening task designed to prevent rehearsal curves show an exponential decay (figure 4.2B) and reach asymptote within 30 seconds , which is the maximum range of STM. For this reason , the correspondingdecay can be fully explained by STM decay . Conversely , residual information held beyond this short-term range can be attributed to an intermediate store independent of either the articulatory es , since the latter are prevented loop or other controlled cognitive process the . Therefore the by , experimentaldesign processresponsiblefor this memory is automatic . The two curves are similar in controls and in unilateral ' temporal lobectomy and Korsakoff s patients. This can be attributed to a close to normal cortical function sustaining the STM capacity , combined with a residual hippocampal function attributed to the spared side of the Hs. Alternatively , it may be attributed to the diffuse and partial lesions of Korsakoff' s syndrome , in a situation where to -be-remembereditems are well within the STM span (three items). This interpretation is strongly suggested by patients with bilateral lesions who exhibit catastrophic forgetting when faced with an interferencesituation. Moreover, electrical brain imaging evidence suggeststhat unilateral hippocampectomycauseslittle changein cognitive ERP patterns (in particular, P300) recorded on the scalp , with the of limited areas of the lobes where a trend possibleexception very temporal toward asymmetry is seen(Johnson 1995 . , ) In the classic free recall , two -component task , lists of unrelated items exceeding the memory span (e.g ., ten words) are presentedto subjectsat a . Subjectsrecall the items after pacesimilar to that of the previous experiment

et al.: Memory , Amnesia , and Banquet

I Interactions

a variable delay, during which they perform an intervening task. The main difference&om the previous task is that the memory load here exceedsthe -recall condition, rehearsalis prevented . In the zero delay, immediate memory span -recallednew stimuli in spite of the of to be by ongoing presentation the absenceof an intervening task. The recall curves (figure 4.2A ) for controls show very robust primacy and recencyeffects(Glanzer & : Amitz , 1966). The recency effect is classicallyattributed to a persistenceof items in STM. As such , it disappearsboth in patients and controls when they perform an intervening task for 30 secondsbetween presentation and recall. The primacy effect is also classicallyattributed to LTM . But in the context of our model, it is dependenton ITM . The primacy effect deterioratesat zero delay even in patients with partial preservation of medial temporal lobe functions (seefigure 4.2A ). It is completely abolishedin the samepatients after 30-second , whereasit persistsin normal controls. Thus, partial preservation delays of medial temporal lobe function is no longer sufficient to prevent intermediate , memory deficiency when the memory load exceedsthe STM span as is the casein the short-term forgetting task when the memory load falls within the STM span . In summary , patients with both bilateral and unilateral lesions of the Hs show close to normal immediate STM for verbal information within the STM span (typically three items) and can extend this temporal range when rehearsalis possible . Both types of patients show catastrophicloss of information when there is overflow beyond the strict STM span . This loss of information is much less systematicand dramatic in normal subjects . There are some indications that the STM span may be reduced in patients. When information is well within the memory span , patients with unilateral temporal -range lesions or Korsakoff' s syndrome show almost normal intermediate . This is not the case for those , independent of rehearsal memory capacity with bilateral temporal lesions . It must be emphasized that the side of the unilateral temporal lobectomy influenceswhich type of performanceis spared(verbal or visual). Thus, both the forgetting curves and the primacyrecencyeffect in partial lesionssupport the distinction betweenan immediate STM and an ITM . Further , the distinctive behavior of patients with bilateral extensive ablation or lesion of Hs corroboratesITM as a possible automatic -based ) for WM . support (non- rehearsal

BrainImaging
The development of brain imaging techniquesbased on electroencephalo ), magnetoencephalography (MEG), positron emission tomoggraphy (EEG ), and functional magnetic resonanceimaging (fMRI ) makes it raphy (PET to possible investigate directly the anatomy and function of memory, both in normal subjectsand in patients. The few resultsobtained by thesetechniques . They lend support to both vary according to the experimental paradigm cortical and hippocampalcomponentsof WM .

Introducti O il to Neural Networb

I Imaging Eledrical Brain


. ERPsare scalppotentials Electricalbrain imaging can be used to record ERPs that reflect different cognitive processingoperations performed by the brain. I'\d top ERP responsesare individualized on the basis of their latency a , such as N 200 ography. Somereflect automatic identification of the stimulus as others such at 200 ms mismatch (negativity ) negativity ; processingneg, , 1982) reflect attended stimulus processing (Banquet ativity (Naatanen : Renault Smith, & , 1990); and still others such as P300 (positivity at 300 ms) reflect context processingor updating. ERP investigations of WM confirm the involvement of various modality-specific or associativecortical regions for componentslater than P300 , suchas P600 and over (Ruchkin et al., 1991, 1992). Surprisingly, P300 , which is related to context updating and probability : Donchin, : Lesevre ,& , 1981; Johnson & , Renault ( processing Banquet " " 1982), does not reflect the cortical explicit component of WM . These results support the involvement of cortical areasin WM , but do not exclude the participation of deeper structures , such as Hs, in typical WM or other electrical . Indeed of , ) predominantly recordings (e.g ., EEG scalp types paradigms , P300, if not exclusively, explore the cortical mantle. Furthermore Wilson in recorded the like activity hasbeen , , hippocampus(Halgren, Squires et al., 1980). But the P300 hippocampal source is not the generator of cortical P300s . Further, it has little influence on cortical P300 generators , since cortical does not induce left or any significant right hippocampectomy . First, far-lateral . There are , however, two notable exceptions asymmetry temporal electrodes(T5 6) in oddball paradigms show a reduced P300 on the side of Hs removal. Second , left hippocampectomyinduces a change in P300 behavior, rather than P300 asymmetry , along with a performance deficit in the number of correctly recognizeditems in recognition paradigms , 1995). involving stimulus familiarity (Johnson
Positron Emission Tomography and Magnetic Resonance Imaging

Neural correlates of verbal WM involving the articulatory loop have been explored by PET measures of regional cerebral blood flow in a task engaging components of the articulatory loop , the phonological store , and the subvocal rehearsal system . Performance on this task was compared to a simpler condition involving only the subvocal rehearsal system (Paulesu et al., 1993 ). This paradigm allowed experimenters to localize the phonological store to ' the left supramarginal gyrus and the subvocal rehearsal system to Broca s area. These results provide strong support for a cortical component of WM , and more specifically , the articulatory loop . Nevertheless , the absence of significant Hs involvement does not argue against the role of Hs in WM since subjects were explicitly instructed to rehearse the to -be- recognized consonants after a two -minute delay period . It must also be noted that the task involved recognition and not recall.

et al.: Memory , Amnesia , and Banquet

Interactions

, Ojemann , Miezin, et aI. (1992) have presented one of the most Squire salient arguments for combined cortex and hippocampus involvement in ITM . In a delayed (three-minute), cued recall paradigm involving no interfering task , subjectslearned visually presentedword lists (fifteen words). During cued recall , PET scans revealed significant activation of the right and hippocampus parahippocampalgyrus, plus the right (and to a lesser degreeleft) frontal lobes (figure 4.3) comparedwith baselineconditions. The left hippocampalregion and amygdala did not change their activation level during cued recall. In our model, this task correspondsto a test of ITM support for WM , because recall was delayed and the memory load exceededthe STM span . The absenceof any interfering task during the delay may have favored rehearsalstrategies and a bias toward cortical activation. But this bias may have been limited by the length of the item list. Nevertheless , the activation of Hs along with the frontal lobes provides strong support for a WM with two components , cortical and hippocampal , combined in aclosed. The selective activation of the right Hs can be explained as a loop system bias toward processing the visual characteristicsof the word forms introduced by the cued recall (visually presentedstem completion). The dual involvement of Hs in both priming and cued recall (see figure 4.3) contrasts with the selective involvement of the frontal lobes in cued recall. This suggeststhat the declarativeaspectof memory involved in recall (in contrast to the automaticity of stem-completion priming ) is a cortical, . At the sametime, it justifies the term possibly frontocortical, characteristic " automatic " component of WM for ITM . Squirehas interpreted theseresults in light of his theory of Hs' s role in the declarativeaspectof memory. Thus, Hs activation in the simple priming task is accountedfor by a covert implication of explicit memory during priming (Squireet aI., 1992). We propose a somewhat different explanation . We attribute the involvement of Hs in -completion priming to the automatic characteristicof this task which stem matches the automatic aspect of ITM support for WM . This explanation predicts a decreasein prefrontal activation to a level close to the baseline condition during the priming task. This is a reasonableprediction if we assumethat the declarativecharacterof memory is independentof the process of consolidation and that the final repository of memoriesis cortical. Recent results from PET paradigms manipulating spatial WM factors in humans show the involvement of the hippocampal formation during these tasks(Owen et aI., 1996; Petrideset aI, 1993). More important, a PET study of prefrontal cortex (Pf )-limbic system interactions during a WM task for faces involving three delays (3.5, 12.5, and 21 seconds ) shows dominant interactions between (1) extrasmate areasand anterior temporal-inferior Pf cortices at no delay, (2) hippocampal and cingulate areasat short and medium , and (3) Pf-cingulate and Pf-temporal/ occipital cortices at long delays delays (McIntosh et al., 1996). Such results are consistentwith our hypothesis . Cortico-limbic concerningthe early involvement of Hs during processing interactions can be interpreted as a processfor maintaining an iconic repre-

Introduction to Neural Networks

) - asu : swa Nowew 181 .

esuodse ~ ~ ~

sentation of faces , supporting our interpretation of the function of the cortico hippocampal loops. Decreasedlimbic involvement at long delays may meanthe end of fast learning (with transition from active potentials to STP ) and transfer of activity to previously involved cortical areastriggering the processof consolidation.

HUMANVS . ANIMALWORKING MEMORY


The WM construct emerged almost simultaneously in animal and human . In parallel with human neuropsychologicalstudies of WM (Badde research ley & Hitch, 1974), the conceptwas developedindependentlyby Olton et al. (1979) in animal learning to refer to the capacity to retain information across trials within a test session . These authors were among the first to propose that the Hs may be necessaryfor animal WM (memory for recent information that has current and specificrelevance ). However, the hippocampusmay also be equippedfor active information selectionand processing , in line with human models of WM . In the radial mazeof Olton et al., p :ach arm is baited with food. On a seriesof trials in the samesession , the animal avoids revisiting an arm from which it has already taken food, suggesting that it retains behavioral events associated with entering eacharm. Thus, WM may encode associated with . As such , it may represent specificepisodes specific maze arms the ITM equivalent of episodic memory. But performancemay also be guided by the stronger relative familiarity of cues related to arms visited most recently. In the 1970s , recordings performed on awake monkeys trained on delay tasks were used to extend the concept of WM . These recordings response showed that someprefrontal cortex neuronswere activated during the delay period between stimulus and response (Fuster & Alexander, 1971; Niki , 1974). These activities were supposed to reflect the cellular expression of mnemonic process es. The evidence in favor of a mnemonic process rather than a motor set or any other activity was made more convincing by the . In an oculomotor delayed discovery of some specificity in the response neurons alter their rate for one or a few , responseparadigm only discharge , thereby demonstrating a kind of memory field (Funahashi , target locations Bruce , & Goldman-Rakic , 1989). This result has been extendedto other brain -to-sample (OMS) areas , the inferotemporal cortex during delayed match tasks in particular (Fuster , 1997 ; Fuster & ] ervey, 1981; Miyashita & Chang, 1988). This form of memory has been called activeby some authors , in contrast to passive(Eichenbaumet al., 1994). Passivememory is characterizedby a reducedresponseto familiar or repeatedstimuli. Someneuronsin the inferior temporal cortex fired much less in response to an immediately repeated stimulus in a serial recognition task (Baylis & Rolls, 1987; Rolls, Baylis , Hasselmo , et al., 1989). Item specificity of the neuronal responsewas also demonstrated(Baylis & Rolls, 1987; Miller , Li, & Desimone , 1993). This para

Introdudiol1 to Neural Networks

doxical responsehas been interpreted as a rapid form of habituation. The of -elicited firing may reflect decreased decrementin stimulus responsiveness cortical neurons to familiar stimuli. In contrast to an active memory, this passiveresponsemay be interferenceresistant and persist through the presentation of intervening mismatch choice cues within the sametrial (Miller et al., 1993). Yet both active and passivememory representationsdisappear between trials, suggesting a system that is reset when the information is no , cumulative decrementof response , a gradual longer relevant. Nevertheless to multiple repetitions of items acrosstesting sessionssupports our hypothesis ; there is a graded transition between ITM support for WM and transient LTM basedon LTP support for permanentLTM consolidation. It should be noted that this decrementin responseto stimulus repetition is dependenton . When the samplestimulus has to an automatic , passivetype of processing be actively maintained for comparison with several test stimuli, stimulus , but an increasedone (Miller response repetition does not induce a decreased . 1994 and Desimone , ) There are striking similarities between electrophysiological recordings in , both electrical recording of ERPson monkeys and brain imaging in humans the scalp and metabolic PET imaging. We mentioned previously the functional . In particular, N200 or mismatch significanceof specific ERP components negativity reflectsautomatic identification of the stimulus in an iconic or echoic memory; others such as processingnegativity represent attended stimulus processing ; and still others such as P300 reflect context processing or updating. When two types of stimuli are sequentially presentedwith different to rare ), the P300 and N200 responses probabilities (rare and frequent and frequent stimuli are very similar at first and then becomesmaller for frequent stimuli and larger for rare stimuli. This evolution in responseamplitude takes place after only a few stimuli (less than ten) for P300, but needs more stimulus presentationsfor N200 (Banquet & Grossberg , 1987). This kind of probability processing is fully automatic and is better reflected by P300 amplitude when the subject is not aware of it (Johnson & Donchin, 1982). The decreasein amplitude of response to frequent stimuli can be compared to the decreasein neuronal response to stimulus repetition in . monkeys , selective , processingnegativity , correspondingto an attended Conversely frequency insensitive to event stimuli remains of some , only filtering preselected i.e. in to a match and shows an increased , ( repetition amplitude response activation ) condition. The sameresults hold for PET imaging. Decreased was found in the occipital cortical areasin responseto items that had been recently presented(Squire et aI. 1992). Squirehas interpreted this as areduction in neural computationsrequired for processingrecently presentedinformation . The similarities between human and animal electrophysiological activities extend to a preparatory set. A contingent negative variation (CNV ) paradigm in humans is a formal analog of experiments used in monkeys to explore WM (Fuster , 1997; Niki , 1974). After a warning stimulus (51) and

. and Carlita -HippocampalInteractions Banquetet al.: Memof)' . Amnesia

a delay, an imperative stimulus (52) commandsthe subject to emit a (usually motor) response . Justas the delay neuron recordingsin monkeys show early, late, and intermediateactivity , the CNV has two distinct componentsin the 51- 52 interval, an early one that is related to 51 processingand a late fronto central component with a ramplike activity preceding52. This late component does not just correspond to a motor set, but also reflects timing and -cognitive set (Ruchkin et al., 1992). perceptual FUNCn ON AL MODEL OF HIPPOCAMPO - CORTICAL RELATIONS We first present the functional model of hippocampo -cortical relations , then in favor of different of the In model. the recapitulatearguments components course of this chapter , argumentsfrom neuropsychology and brain imaging have been provided to argue that the Hs is involved in every phaseof information , from the very short-term to the transient long-term. processing " There are also some arguments against Hs involvement in " permanent LTM , although opinions diverge on this issue . Thus we have chosento contrast hippocampalfunction that is devoted specifically to information selection and consolidation with cortical function that supports a direct dialectic confrontation betweenSTM and LTM during learning, recall and recognition. The first claim concernsthe complementarity between cortical and hippo. The cortex is primarily involved with short-term campal memory systems and long term permanent memory registers , even though range variations exist from primary areasto secondaryand associationareas(Lii , Williamson, & : Kaufman , 1992). And it has a relatively specificcapacity for slow learning. The hippocampusis richly endowed with a full spectrumof temporal ranges , from the short-term to the transient long-term, possibly at the exclusion of the permanent long-term. More specifically , this combination of diverse with diverse closed of , , including one memory registers loops different sizes to at least five synaptic relays , seemsunique in the brain. It is probably responsible for the one-exposurelearning that is peculiar to the hippocampus . This capacity constitutes a turning point in phylogenesis because it can be , linked to the development of an actual ontological dimension in individuals, and more specificallyto personality in primates . On the basisof logical arguments , we suggest that fast-transient yet to be experimentally confirmed learning should take place not only inside the hippocampusbut also at the interface of the convergent cortico-hippocampal pathway (either entorhinal cortex or dentate gyrus) and at the interface of the divergent hippocampo cortical pathway (possibly in superficialcortical layers). Conversely, experimentally confirmed slow-permanent learning in the neocortex may be restricted to the level of polysynaptic corti co-cortical connections . The second claim concerns the complementarity of the two aspects of hippocampal transient memory: ITM , which supports WM function, and transient LTM , which supports the process of LTM consolidation. These complementaryfunctions cooperateto fulfill the contradictory constraintsof

Introduction to Neural Networks

storing as much relevant information as possible in a large but limited. The WM function thought to operate both at cortical and capacity system levels is devoted to the segregation between information hippocampal " " be to of its relevanceto survival or its worthy permanently stored because human social relevanceto the personal history of the subject , and information that can be forgotten without major damage . For this purpose , several hard-wired devices implement different functions in various subsystemsof the hippocampus : severe61teringon the basisof stimulus intensity, duration, or repetition (mostly at the level of the entorhinal cortex); orthogonalization of noncorrelated patterns of information, with lumping correlated information or suppressionof redundant patterns (possibly in the dentate gyrus), or both; and correlation of these orthogonal patterns on the basis of spatio) or sequential ordering temporal criteria of co-occurrence(autocorrelation (heterocorrelation ). The CA3 subsystem is specifically equipped for these correlational process es. There is some evidence for temporal order processing of events at the level of CA1 (Granger et al., 1994). All of these functions argue in favor of a specific processing capacity devoted to the Hs, based on ITM . This capacity may be devoted more to temporal aspectsof information (like temporal order and timing ) than to , spatial spatial mapping, which is the focus of current investigation. Specifically a in would include dimension the ( strong temporal mapping sequential ), where temporal order is no longer relevant. Correlational recording of snapshots processing may be the hallmark of Hs and may account for many of , including temporal processing(Banquet aspects its cognitive processing b . the basis of the et al., 1997a On CA3 associativearchitecture , ) , artificial systemscan be designed so that temporal order processingbecomesa special caseof correlational processing . The main differenceis that events occur in sequence . , insteadof taking place simultaneously The other class of transient memory is transient LTM based on LTP, which supports permanentLTM consolidation. The most prominent sites of LTP in the hippocampusare at the CA1 level but they are also in the entorhinal cortex, dentate gyrus, and CAJ . LTP is not specific to the hippo. Different cortical areasare also prone to LTP. What seemsspecific campus to the Hs is the conjunction of systemssusceptibleto LTP with shorter versions of synaptic facilitation. This occurs in an architectureadaptedto facilitate a smooth transition between short-term and long-term facilitation, with no need to resort to externally dependent repetitions of activation. These . two functions of ITM and transient LTM are therefore complementary Indeed , information has to proceedthrough the cortical or hippocampalWM in order to be certified as relevant to the history of the individual. This certification could correspond to the transition from ITM supported by shortduration synaptic facilitation (STP ) to transient LTM supported by LTP. The third claim concernsthe unique mechanismthat supports WM operation and LTM consolidation. This unique process takes advantage of the reciprocal , topographically organized connections between cortex and

et al.: Memory - Hippocampal Interactions , andCarlita , Amnesia Banquet

, combined with the endogenousproperty of pyramidal neurons hippocampus (mostly in the CA3 region) to dischargeby bursts (Buzsaki , 1989), and to synchronize within populations connectedby previously facilitated synapses . Sharpwaves consistently recorded in rodents have also been found in . They may form the basis for reactivation of recently primates and humans facilitated neuronal populations at the hippocampal level, at the cortical level, or at both. They would correspondto what we call the print mode of the Hs. In episodic learning, the print mode would subserve the function attributed to practiceduring the formation of proceduralmemories . Two generalmechanisms in the reactivation of cortical , cooperate patterns triggered by these endogenoushippocampal bursting and synchronization capacitiesor by hippocampalreactions to cortical activation of cue patterns. First, there is reverberating activity between reciprocally connected networks . This has been used in neural network modeling as a mechanismfor , 1976a , b). synaptic weight modification leading to classlearning (Grossberg Second of reverberantactivity , , neural population synchrony, a consequence has been presentedas a plausible temporal coding used locally by the brain for the coalescence of features into a unified percept (Gray, Konig, Engel , et al., 1989; Singer , 1991). The main contention of our model is that the central location of the hippocampusas a site for both convergencefrom, and divergence toward, the different cortical areas allows the Hs to act as an information-selectivepacemaker , 1983; Banquet& Contreras-Vidal, (Banquet 1994). As such it can , selectively synchronizedistant cortical areasthat have beencoactivatedpreviously. This mechanism refresh es memoriesduring WM operation and reactivatescortical patterns during the more lengthy process of LTM consolidation . In the latter case , it createsconditions for the slow facilitation of distant polysynaptic cortico-cortical connections that eventually make recall and recognition independentof hippocampalfunction. The last claim is largely but not universally accepted . Information that is the not be processedthrough hippocampusmay permanently stored there, at least in humans . In addition to anatomical arguments like the relatively limited capacity of the system , the strongest support for this claim comes from the limited range of the ~ trograde amnesiain hippocampal lesions . Most memoriesfrom the distant past are spared . This suggeststhat information is transferredto cortical or cerebellarareas , which are among the most sites of . prominent permanentstorage We now summarizethe argumentsthat support our model. In favor of an , neurophysiological evidence early Hs involvement in information processing that the in Hs activation in suggests lag responseto an external input, to cortical areas does not exceed one or two hundred , compared primary milliseconds . In particular, the P300 ERP component can be simultaneously recordedat cortical and hippocampalsites (Halgren et al., 1980). At abehavLoral level, neuropsychologicalevidence confirms that patients with partial Hs lesions show complete disruption of STM span performancewhen the

100

Introduction to Neural Networks

. The prorogation by memory load is at or exceedsthe limit of the STM span rehearsalof STM in a primary memory (James , 1890) is possible for patients in very limited conditions of memory load, absence of interference , and type of material to be memorized(easily verbalized ). Thus, experimentalevidence supports the existence of a period, starting as early as during immediate STM, when cortical and hippocampal activation and processing overlap. During this period, the cortex can both hold and process information, in , very restrictive conditions. This interpretation explains why , at first glance STM appearsto be Unaffected lesions . But the of by hippocampal importance hippocampalfunction, particularly in WM , may have been underestimatedin . previous investigations We will not summarizeargumentsin favor of a transient LTM in the hip. This would amount to describing again those results that support pocampus a graded retrograde amnesiabuttressing the LTM consolidation hypothesis. At the neurophysiological level, there is an extensive literature on hippocampalLTP and its potential role in learning; however, a review of this literature is beyond the scope of this chapter . Thus, we justify the distinction between STM and ITM , and also between ITM and transient LTM , on functional grounds. Both transient LTM and ITM refer to transiently stored information. The distinction between the two constructsinvolves the behavlorally justified dichotomy between the two functions of WM and LTM consolidation . It emphasizes two points: (1) the LTM consolidation function concerns information that has already been selected for storage ; and (2) therefore it should be supported by more elaborate neurophysiological mechanisms . stated , STM dealswith the immediateSTM span , and stored information Briefly there decays rapidly if it is not rehearsed . There is a floor effect to this decay . This constitutes , and asymptote is reachedin lessthan 30 seconds the maximal range of the STM store. The spared information beyond this ) store. ITM functions lapse of time can be attributed to an ITM (minutes range to retain information as an automatic support to WM . It also supports an important processingfunction; in cooperation with cortical processing , it allows segregationof informaitioninto storable and nonstorableparts. At the neuropsychologicallevel, there are several arguments in favor of ITM . First, a primacy effect in the dual-task , free-recall paradigm , and its disappearance in patients with partial Hs lesions , suggestsa short-lived ITM located in Hs. The classic interpretation of the primacy effect as an LTM effect is no longer tenable , in view of evidence for a consolidation phase " " LTM . Further, the delay between first-item presentation preceding permanent and recall is on the order of a few minutes maximum. This time range has been characterizedby researchon amnesiaas pertaining to an intermediate store. Second , the chance level performance of hippocampal on various delayed recognition tasks(face recognition in particular), patients whenever an intervening task is interpolated between performance and recall , and their failure to perform (or to learn by practice ) short versions of a

-Hippocampal et aI .: Memory . andCortico Interadions , Amnesia Banquet

visual mazesupport a deficit in ITM . This deficit is not just a memory deficit. It is a cognitive deficit involving the inability to partition a complex goal into accessible of damageto the frontal subgoals(this is also a consequence lobes whose function relies on WM process es). ITM can be viewed as an -responsetasksin animalsand automaticaspectof WM . Many of the delayed -to -sample tasks humans , particularly delayed matching- or nonmatching , ITM . show a dramatic with extensive Hs lesions . They explore perturbation This perturbation in humans is conditional on the prevention of rehearsal and on the nature (geometric ) of the to -be-memorizedmaterial. Neurophysiological evidencein favor of ITM is lessfamiliar, and therefore little emphasized in the literature. First, the architecturalconcatenationof Hs and of Hs and cortex in loops with variable numbersof synaptic , subsystems was noted , . Yet this relays long ago, in particular the classictrisynapticloop information has had little impact on models of hippocampal function. Furthermore , the crucial importance of the trisynaptic loop as the information highway through the hippocampus has been deemphasizedby the more recentdiscovery of the functional importanceof shortcutsfrom the perforant path through CA3, CA1, and subiculum (Amaral & Witter , 1989; Jones , 1993). There is evidence for the actual operation of these loops, at least in experimental conditions. The sameCA population of neurons first give an electrophysiological responseto a single stimulation of the perforant pathway , and delayed attenuatedsecondand eventually third iterative responses (Buzsaki 1989). The mechanismresponsiblefor the topographic stability of the activated neuronal population during iteration is unknown. In neural network modeling, recycling information through a closed loop is a classic design for maintaining information in STM. In physiological situations , the maintenanceof relevant memories in an active state does not need to be , only punctual. It may permanentfor the entire duration of task performance have a twofold functionality. Cortico-hippocampal loops may maintain memory in an active state punctually during the phasesof processing that actually use this information. In this way, theseloops originate the activation of cortical (frontal or temporal) neurons that fire in the interval between two stimuli or between stimulus and response . Also, the intrahippocampalloops a responding by few iterative electrical responsesto a single perforant path excitation may foster STP in the corresponding circuit relays and make a transition with LTP. Relative to this last point, almost every hippocampalsubsystemseemsto demonstratea capacity for LTP. In spite of this apparent uniformity in the LTP process , different phenomenaseem to correspond to different mechanisms . Post-tetanic potentiation (PTP) does not exceed a few minutes ; STP lasts for about one hour; LTP itself can be classified into three groups : LTPl lasting hours, LTP2 for a few days, according to its rate of decrement and LTP3 for weeks or months. The samesynapsesappearable to undergo , according to the intensity or duration of anyone of these transformations the activation, at least at somehippocampalsites. They may represent , under

Introduction to Neural Networks

natural conditions, different steps of synaptic facilitation, generated or not according to the intensity or duration of the eliciting event. The alternative exhibit specifictypes of hypothesis is that different hippocampalsubsystems LTP, according to the widely distinct functions they appear to perform. In this regard, entorhinal cortex has the capacity both for classicLTP and also for very short facilitation (during an input with increasedfrequency ) that allows action potentials to overcome the hyperpolarization barrier (Jones , 1993). Dentate gyrus seemsto be endowed with a wide variety of ranges , and may favor cooperativity in its mode of LTP generation becauseof the high convergence of perforant path fibers that it receives and its high threshold for activation. The facilitation of the mossy fiber- CA3 synapse needspostsynaptic depolarization for its generation , even though the induction of LTP there does not depend on NMDA receptors , but rather on opiate which are recruited activation. , by by repetitive Conversely receptors , the CA3 systemmust be more prone to associativity. It hasboth ITP design and LTP, according to its synaptic contact type (mossy fiber vs. associative commissural ). CA1 LTP constitutes the referencemodel of synaptic facilitation whose generation seemseasier than at the CA3 level. Some forms in specificconditions (Staubli & : Lynch, 1987). may even be nondecremental In conclusion , a whole temporal spectrumof synaptic facilitations is available in the Hs. It is prematureto claim that the hippocampusis the only system to be so richly endowed . It may also be true of cortical potentiation. The consolidation hypothesis requiresa precisemapping between specifichippo, particularly at the level of campal and corresponding cortical populations -cortical connections the hippocampo . There is no anatomical support for such precise hard-wired mapping, in spite of evidencefor a loose preservation of topology between cortex and hippocampus in the longitudinal dimension . A plausibleimplementation of suchmapping may involve a transient LTP at the level of the reafferentconnectionstaking place at the interface between hippocampusand cortical layers (Burnod, 1990). LTP is known to take place in different cortical areas . There may also be a gradient in the duration and diversity of the memory registers on the path from hippo. In any case , we can system campusto cortex through the parahippocampal state that a unique multiple-loop system resides in the Hs whose synaptic relays may be facilitated for a wide variety of durations. Topographically specific facilitations are more or less lasting according to the conditions of their induction or location, or both. The progressiveelaboration of information , from entry to exit, along the transversaldimension of the hippocampus have some value for heuristic may neurobiological investigations of the connectivity inside the hippocampus . There may be some gradient along this dimension in LTP generation , in LTP duration, or in both. There is already somepreliminary evidencefor this hypothesis . As already mentioned , this continuum of memory registers that sharethe characteristics of lasting beyond the short range and still being transient can be classifiedinto short-lived ITM and transient LTM . These support, respec tively , the functions of WM and LTM consolidation.

-Hippocampai et al.: Memory Interactions , Amnesia , andCortico Banquet

MATH EMA DCAL MODEL


We have seen how Hs function is based on a full temporal spectrum of memory registerswith different lifetimes. There are some experimentalindications that Hs is involved in processingthe temporal order of events at a basic level, and also in timing . The encoding of temporal order by differential synaptic facilitation, according to the order of stimulation, may take place at the level of synaptic contactsbetween Schaffercollateralsand CA 1 pyramids (Granger et al., 1994). Theseexperimentalresults confirm an algebraic model of STM and LTM for temporal sequences formulated by Gross 1978 . The network model of temporal sequence ) berg ( learning we present here is basedon this algebraicformulation, and involves STM and WM . As , it is possible to implement temporal order basedon ITM suggestedabove and autocorrelation properties of the CA3 field. Temporal order can also be : Levine, 1994; Grossberg , 1978; Reiss preservedand learned in LTM (Bapi & & : Taylor, 1991). The hippocampushas also been implicated in timing behavioral responses : Thompson (Berger & , 1978; Solomon , 1980). Neural mechanisms underlying the hippocampal adaptive timing function during conditioning were proposed by Grossbergand Merrill (1992), among others. This basic computa tional competencemay also be shared by the cerebellum(Bullock, Fiala : ,& 1994 & : Keele 1989 . The model , ; here , is based on ) Ivry Grossberg presented the sameprinciples of population learning and coding of time by a limited assembly of neurons whose dynamical range of activation varies along a biologically plausiblecontinuum. This timing system is integrated in a more , including temporal order processing (figure complete processing system 4.4). The network is designedto emulatethe electrical responseobtained on the scalpin humans , in a temporal conditioning paradigm . The paradigm can be viewed as a formal analog of a CNV paradigm . Once interval learning has taken place , the timing network gates the main processing trend that involves sequence and probability learning. In our model, relations among Hs and other structures , particularly frontal cortex for temporal order processingand cerebellumfor timing , are viewed as specialcases of the more global function of Hs as a rehearsal systemfor the of information. This ( ) rapid acquisition temporal interpretation is supported the fact that this of information by , particularly time intervals, learning type is not really suppressed . It just requiresmore trials to by hippocampallesions be learned , probably by the cerebellum . , in conditioning paradigms

TEMPORAL ORDER NETWORK


Different models of temporal order processing have been designed . The most recent ones emulate temporal processingattributed to prefrontal cortex . Guigon, Dorizzi, Burnod, et al. (1995) have designed a model directly . Temporal inspired by electrophysiological studies performed in monkeys

104

Introduction to Neural Networks

Choice Chunk ' "


Timing

Probability / "

.
f
--- --Categories

Temporal order

-modulesystem order for timing of the multiple , , andtemporal Figure 4.4 Top: Architecture the : Choice to three different ; whim sends systems layer original sequence performs outputs the repetitive eventlocalandglobalfrequency ; Chunklayerlearns Probability layerintegrates of the . (Right of a probabilitycondition ) Neuralnetworkcomponents prototypicalsequences and the : is of three module . Tom order Nonnalization , , (Tom ) layersRepeat composed temporal three different . It sends to from the Item . It receives Tom bufferlayer systems layer outputs inputs a population in Agure4.4. (Ltft) Timing circuit representing ; Only probabilityis represented . In ; Drive , output ; P, now print signal , inhibitoryinterneuron codingof the timeintervals . of expected timing response

are learned , in their model, by processing units that commute sequences between two stable statesof activity (bistableunits) in responseto synaptic activations. After learning , the sustainedactivation of a given neuron represents the selective memorization of a past event, selective anticipation of a . Thus, the model reproduces future event, or the prediction of a reinforcement the functions of neuronsencounteredin frontal cortex. Bapi and Levine (1994), Levine and Parks (1992) and Parks and colleagues (1992) have designed network models of frontal cortex function. Sequencelearning is in LTM securedby storing the transitions between the events of a sequence are encoded in acom (synaptic weights). The different learned sequences . These models are similar in some pressed form and can be categorized WM for of our own model to temporal order and probability coding respects : Contreras Vidal, 1992a , b, 1994). Levine and his , b, 1993a (Banquet &

- Hippocampa , and Cortico J Interadions Banquetet aI.: Memory, Amnesia

Categories Timing

Items

' models andour modelareinspired by the same colleagues designprinciples , 1978 , Changeux , andNadal(1987 (Grossberg ). Dehaene ) haveimplemented a networkthat canlearntemporalsequences based on biologicalproperties of allosteric . receptors
Temporal Order as a Spatially Distributed Amplitude Gradient Grossberg (1978) first proposed an algebraic STM model that codes temporal order by amplitude gradients of node activation (representing neural ). Thus, spatially encoded event categoriesare populations encoding events modulated by amplitude gradients that encode temporal order. A temporal pattern is encodedas an amplitude-modulated spatial pattern with well-char acterizedprocessingand learning mechanisms . Yet a free-recall , dual task demonstrates , recencyeffects , or primacy effects both according to the length of the sequenceto be recalled (Atkinson & : Shiffrin, 1968). In order to comply with these psychological constraints , Grossbergdesignedtwo principles: 1. An invarianceprinciple , whereby amplitude-modulated spatial patterns of activation acrossnodes in STM are stored and reset in responseto successive events in such a way as to leave unchangedSTM temporal order codes of past events . The corresponding algebraic implementation involves the uniform multiplication of a previous pattern t - 1 by a factor omega (i ) as the ith item is instated . 1976a 1978b , (Grossberg ) present 2. A pseudonormalization rule to , designed implement the limited STM capacity in a more flexible way than straightforward normalization. By this rule, S(i ), the total activity at time i, tends asymptotically to S, acharacteristic of STM limited capacity . , as i increases Grossberg(1978) proved that the rate at which S(i ) tends to its asymptoteS determinesthe form of the STM activation patterns : . Primacy : x (k - 1) > x (k ) for all k = i . Recency : x (k - 1) < x (k) . Bowing: combinesprimacy for early items and recencyfor later ones Bradski , Carpenter , and Grossberg(1992) designeda shunting architecture that realized invariance and pseudonormalizationas emergent properties of the system dynamics . This basic architecture was adapted here in order to insert it into a system capableof coping with a continuous flow of inputs from two or more distinct categories . Theseinputs could be, for example , the of one or several modality outputs specificunsupervisedcategorization networks like adaptive resonancetheory (ART ) working in parallel (Carpenter & : Grossberg : Reynolds , 1987a , 1987b , 1990; Carpenter , Grossberg ,& , 1991). Most of the equations for the different stages of the temporal order network (TOM ) are variants of typical shunting network equations . These

106

Introduction to Neural Networks

equations can be viewed as adaptations of the classic Hodgkin -Huxley single - neuron equation to neural populations . As such, they comprise three basic components : an excitatory term , an inhibitory term , and a passive decay term . The first two terms include a multiplication factor that represents an excitatory (or inhibitory ) saturation point corresponding to excitatory (or inhibitory ) equilibrium points in the Hodgkin -Huxley equations .

TOM Architecture and Simulation


The categorical (amodal or supramodal ) inputs in our TOM model first go a module that sorts out through repeat repeateditems belonging to the same in a the size of a STM . Then the sorted items enter category sequence span a WM module for sequenceprocessing that has the typical STM span (approximately seven items), and correspondsto a cortical memory register. It contains a layer that remains active in the interstimulus interval (ISI) between two successive items and can be comparedto cortical neurons that remain active in the interval between two stimuli, as documentedin the inferior temporal cortex. The first of two layers in the WM module is anormal ization layer, active during input time. It is reciprocally connected with a buffer layer, or temporal order memory proper, that tracks the activity of the normalization layer in the ISI between two stimuli. A flip -flop pattern of activation is instantiatedbetween normalization and temporal order memory layers. By simple variation of the passive decay coefficient in the normalization of items the size of an STM span can be encoded , sequences layer equation by the system with either a primacy gradient, a recency gradient, or a combination of the two , in accordance with experimental results . P300 variations in humans encode the order of the events amplitude temporal Smith & Giinther 1992 al. et . 1976 These , results motivated (Banquet , ; Squires , ) our model. In our experimentalparadigm(Banquet& Grossberg , 1987), of about 500 stimuli were delivered to subjects in blocks of 80. sequences Feeding such a sequenceto the WM network when it was operating in a recencymode gave rise to a traveling window of waxing and waning activity in the two WM layers. Furthermore , presenting the system with sucha long sequence gave rise to in the that receive emergent properties upper layers temporal information from TOM , namely, a chunk layer for categorization of the sequences and a probability layer for learning the probability rule; patterns of probability layer activation are shown in figure 4.5. In particular, the activity in this last , since it tended to an equilibrium layer behaved as a P300 ERP component level encoding both prior probability as a baseline and temporal order of events and local probability in the fluctuations of this baseline . This combined encoding in the model is dependenton the integrity of both STM and ITM attributed to the hippocampus .

107

, and Cortico-Hippocarnpallnteradions Banquetet al.: Memory, Amnesia

0 .8 0 . 6 OA 0 .2

0 . 8 0 6 . 0 . 4 0 . 2 0

f
g ~ C . g c 0 0 0 8 . .

0 G 0 -

TIMING
There are several scales of biological and mental time . The scale evaluated here is relevant to coordinating integrated systemic processes as the basis of behavior . Life situations in which timing and temporal ordering of events cooperate are common , not only in perception and cognition but also in motor control . This is also true of various artificial systems devoted to sequential pattern recognition , such as speech and writing , or motion pattern generation , as in robotics . There is a strong interdependence between the different types of temporal processes. In particular , the brain ' s ability to learn sequences and probability is strictly dependent on the duration of the interval between events.

Timing and ClassicalConditioning


Grossberg and Merrill (1992) and Bullock et al. (1994) have integrated a spectral timer (basedon an array of cell populations responding at different rates ) into a classicalconditioning circuit. They were able to explain data ' concerning an animals ability to accurately time the delay of a goal object, based on its previous experience in a given situation. This timing ability controls the right balancebetween exploratory and consummatory behaviors ' , and seemsto be essentialto the animals survival.

Timing, TemporalConditioning , and Cognitive - Motor

Process es

The paradigm that we used in humansto explore cognitive-motor process es : Grossberg (Banquet & , 1987) is a formal analog of temporal conditioning in animals . There is no distinct CS. The equivalent of the US is presentedin Bernoulli seriesat regular intervals. After some practice , the time intervals )
) Activation of the probability layer for two categoriesA and B of 600 events Figure 4.5 (Top (500 in b and f ). In the left and middle columns P(A ) = .2, P(B) = .8. In the right column one or two probability shifts occur at times indicated by arrows, from .2/ .8 to .5/ .5, or reverse . The left column presentslinear (a) or sigmoidal (b,c,d) feedbackfunctions. Furthermore , from b to d probability node dynamics becomefaster . Then the system presentslarger osdllations in to sequentialdependencies and local probability fluctuations. The experimentalrelation response -state amplitude is better fitted by d. The middle column between local fluctuations and steady has the sameparameters as the left one, except that LTM weights are absent , suppressingLTM learning. Oscillations and asymptotic value still exist, but amplitude probability coding is biased toward intermediatevalues . The right column featuresresponsesto probability changesat different (arrows): after 400 stimuli in i and k. and after 200 and 400 in j and points of the sequence I. The system adaptability to new conditions depends on previous sequenceduration and on with faster dynamics , faster tracking associated . (Below probability node dynamics ): The module for the evaluation of the interstimulus interval learns the time interval between stimuli in a of events . After learning sequence , the system becomescapableof antidpation of a response even in the caseof an omission in the input sequence . The peak responsecorrespondsto the time of presentationof the stimulus being attended to.

109

-Hippocampal , andCortico Interactions Banquetet al.: Memory, Amnesia

is learned , and a responserelated to priming and preparation starts prior to the onset of the stimulus. Thus, the factors that were manipulated involve es. Then, it becomes learning a time interval and cognitive and motor process possible to addressthe problem of timing in the context of a complex task es , cognitive, and motor process , embeddedin a learning involving perceptual . We the role of in the paradigm timing emphasize planning of a perceptual and motor set to the event-stimulus , cognitive, (prior preparatory occurrence ) and also in the programming, coordination, and integration of these multimodal process es after stimulus delivery. Indeed , any cognitive in is since motor and , cognitive properformance part irreducibly sequential of events . But we cesses on evaluation contend that dependpartly perceptual a non negligible part of learning fosters parallel rather than sequentialprocesses , cognitive, and motor modalities. This claim is consistent among perceptual with the shift from controlled serial to automatic parallel process es task . during learning
Spectral

Timing Network : Design Principles

Extra constraintsmust be brought to bear on a model designedto implement . One of the most conspicuousconstraints complex cognitive-motor behaviors derives from the capability of the brain to processan enduring flow of information , sequentiallydelivered at variable speedsand loads. The brain is able to extract and learn latent regularities in the midst of noise and irrelevant variations. Then it can perform a specific adaptive response , or initiate an integrated global behavior. Several design problems had to be solved in the network that emulates these process es: (1) the significance of each stimulus , and its bivalence as both a CS and a US suggestparallel pathways to the spectralcell population; (2) the continuous flow of inputs forbids any specificcoupling by pairs, since any stimulus belongs simultaneously to two pairs, one with the preceding and one with the following stimulus ; (3) the continuous flow of inputs does not allow the system to be reset after each trial, as could be done after pair . The reset had to be built into the system . In spite of these presentations constraints , the design principle of spectraltiming basedon population coding is robust enough to support a system that functions correctly even under severeconditions (Banquet , Gaussier , Dreher, et al., 1997b).
' System s Architecture and Simulations

The spectral timer (see figure 4.4) is made of a battery of thirty nodes (empty circles ). Eachnode is endowed with its own dynamical range. Thus, the entire population cooperates to cover the whole plausible biological range. The spectraldecompositionand learning of time intervals by the node populations is made possible by a double gating of these node activations: first, by a depletable neurotransmitter gate equivalent to an ITM (filled

110

Introduction to Neural Networks

), and second , by LTM weights (filled half circles squares ) that sample and learn the activation spectrum corresponding to a particular time interval. Learning is modulated by a transiently active print (P) signal, turned on by sufficiently large and rapid increments of the input activation. The transformation of the input into a sharp phasic activity results from interaction with a feedforward activated inhibitory interneuron, In. This feature is a . widespreadarchitecturaldesign presentboth in Hs and cerebellum The global responseof the system (seefigure 4.5) comesfrom integrating the activities of all doubly gated activation signals . This timing drive serves to gate the activation of the other structuresof the architecture , such as the and the . categorizer sequencingsystem

DISCUSSION
There are two dimensions in our model, temporal and spatial , that correspond to the transverseand longitudinal axes of the hippocampus . In this , we emphasizedthe temporal aspectof hippocampalprocessing . We chapter and contrasted the neocortex and . The neocortex is compared hippocampus endowed with dual short term and term mostly long capacitiesrelying on different functional aspectsof the sameneural networks. But it is also plausibly endowed with the capacity for fast-transient learning at the synaptic interfacebetween hippocampusand cortex. The hippocampushasa full array of continuously varying memory ranges , coupled with a system of loops, both intrahippocampal and between hippocampus and cortex. The former . The appear to be devoted to the iteration of single-event representations , the array of registers and complementary operation of the two systems loops, along with the event correlation capacity of some of the hippocampal networks, may account for the capacity to register episodic and factual memories . For the sake of clarity, we have partitioned this continuum of , which correspond to the functional distinction temporal ranges into two classes between WM and transient LTM . WM includes , but is not restricted " LTM . to, STM. Transient LTM is the basisof a " permanent We emphasize how thesetwo functions are complementary . WM , by its processingcapacity , should be critical to the selection , elaboration , and organization of information flow in order to decide what should be permanently stored in LTM . Transient LTM would be the intermediary step between selected information and permanent store. Last but not least , the actual neurobiological mechanisms responsiblefor WM support and LTM consolidation should be similar or even identical. Both process es should be basedon reactivation or reenactmentof cortical patterns by endogenousand synchronousactivation in the hippocampusof recently coactivated neuronalassemblies . , a loose topography is Along the longitudinal axis of the hippocampus , at the level of the preservedbetween cortical and hippocampalconnections . This allows for multiple possibilities anatomy rough topological correspondence with respect to the precise functional mapping between cortical

, and Banquetet al.: Memory, Amnesia

Interactions

networks and the hippocampus . This precisemapping would be functionally determined by previously hypothesized fast-learned transient and topologi cally specificsynaptic facilitations at the interfacebetween hippocampusand cortex. The mapping referred to here does not directly concern the external space(cortical areassuch as parietal cortex seem more appropriate for the . Further ), but the cortical space long-term storage of detailed spatial maps should tell us how the mix modeling hippocampusmanagesto up all cortical areasfor the sakeof correlation at the CA3 level, and simultaneouslyrecover topographic information in its output for the needs of specifically targeted consolidation of cortical LTM . Direct connectionsfrom EC to CAt may play a crucial role in this restoration of topology . Our model integrates different elements , some of them already parts of other models . Other models relating to hippocampalfunction can be loosely classifiedaccording to their emphasison explaining anterogradeamnesiaand , or both. Unfortunately, due hippocampalprocessingor retrograde amnesia to spacelimitations, we cannot review thesemodels here.

CONCLUSION

Cortical processing involves two memory registers , STM and permanent LTM . The transition from the long-term to the short-term seemsto be direct and normally encountersfew problems . As such , it is plausible that the ana tomical substrates are topographically identical (STM representing active forms of LTM ) at a gross level of analysis , even if the supporting neurophysiological mechanismsare different at the fine level, involving either -chemical changes electrical or durable structural . Conversely , the transition from STM to LTM store follows a more intricate path, probably for the sake of optimizing the amount of information stored, but also for securing the , which constitute the unique history of eachliving storage of unique events being. Betweenthesetwo extreme ranges(STM and LTM ), only minor variations from primary to associativeareascan be recorded , with a tendency for the temporal range of memory to increaseas a function of the complexity of . processingperformed by theseareas The specificity and vantage point of Hs concerns both topographic and . The topographic aspect of Hs specificity as a unique temporal dimensions site of compact input convergenceand output divergence has been extensively . It has been credited with the correlational function of Hs. emphasized This anatomicalcharacteristicis complementedby a loose topological correspondence between cortical and hippocampal systems in the longitudinal dimension . This loose correspondence may be transformed into a dynamic learning dependentprecisemapping between neuronal populations to implement the topologically specificconsolidation function. This function may be implementedvia the fast-transient learning capacitiespresent within the hip. pocampusitself, and also at the interfacesbetween cortex and hippocampus The emphasisplaced on the spatial function of Hs was detrimental to the

112

Introduction to Neural Networks

-important temporal function. This function results exploration of the no-less from the capacity of Hs to play very flexibly on a whole spectrumof registers from the short-term to the long-term, but possibly excluding permanent LTM . The unique characteristicof Hs would be the conjunction of this array of registers with a wide variety of variously sized loops. These complementary " " hardware constraintsbetween cortex and hippocampusdetermine the type of cooperation established between the two structures . Further results from the direct contact of the cortex with the environment complementarity , whereasHs is the only structure so easily prone to autoactivation. That leads , in the pathological domain, to seizureactivity . The hippocampushas the structural and biological equipment to cover the entire memory spectrumfrom STM to LTM . Nevertheless , at the behavioral level, this continuum can be split into two types of transient memoriessubserving two distinct functions. Short-lived ITM , basedon STP , supports an automatic component of WM . Transient LTM , based on LTP, supports the LTM consolidation function attributed to Hs. The two functions of WM and LTM consolidation are complementaryand nicely articulated . Only information in WM canaspireto permanentstorage previously filtered and processed by entering the transient LTM store for consolidation. A unique mechanism sustainsthesetwo aspectsof intermediatememory: namely, resonantactivity " in reciprocally connected networks, which can be viewed as a " nonlinear expression of reverberating activity in closed feedbackloops. This mechanism , or at least fixed temporal relation, provides for synchronization between activities in distinct neural populations, which is supposed to be for the Hebbian type of synaptic learning. necessary The cortical controlled component of WM involves corti co-cortical loops. Someof thesecortical loops, suchas the articulatory loop, may be specificto the human brain. The automatic component of WM involves hippocampo -frontal ones cortical loops, particularly hippocampo . This reverberating loop , specifically and punctually involved in WM , is replaced for longer system delaysby synaptic facilitation of various durations, as documentedin the Hs. These synaptic process es sustain transient LTM , which operates the consolidation of permanentLTMs at the cortical level. Thus, with the exception of cortical STM, the transienceof storageat variable degreesis a hallmark of -cortex interfaces es and hippocampus . This transience hippocampal process endows the system with permanentavailability for the processingand transient storageof new information. ACKNOWLEDGMENT We are indebted to ElaineGaribobo for useful commentson the manuscript . REFERENCES .. & t Squire . (1994 consolidation and themedial lobe : A simple Alvarez ,P . L. R ). Memory temporal - 7045 network . Proceedings model the . . 91 . 7041 of National Academy ofSciences

113

I Int~radiQn~ , and Cortico- Hippocarnpa Banquetet aI.: Memory, Amnesia

-dimensional of the hippocampal Amaral : Witter, M. P. (1989 , D. G., & ). The three organization - 591 data . Neuroscience . : A reviewof anatomical fonnation , 31 , 571 : A proposed andits control . M. (1968 Atkinson : Shiffrin , R , R. C., & ). Human memory system : .), Thepsychology and motivation . In K . W. Spence & : J. T. Spence es (Eds of learning process . : Academic Press Advances in research andtheory , Vol. 2 (pp. 89- 185 ). NewYork . Oxford . , UK: OxfordUniversityPress , A. D. (1986 ). Working memory Baddeley .), Thecognitive neurosciences . In MS . Gazzaniga , A D., (1995 (Ed ). Workingmemory Baddeley - 764 . : MIT Press , MA (pp. 755 ). Cambridge .), The . In G. A Bower(Ed : Hitch , G. J. (1974 , A D., & ). Workingmemory of psychology Baddeley in and . 47 . New York: Academic and : Advances research Vol. 8. 90 motivation , ) ( theory learning pp Press . - and short and the distinction between : Warrington , (1970 , A. D., & ). Amnesia Baddeley long and Verbal Behavior 9 176 . . Verbal 189 termmemoryJournal , , of Ltaming in man . . intra of the EEGduringsleep . P. Inter and , J (1983 ) hemispheric relationships Banquet 55 and 51 59. Clinical , , Neurophysiology Electroencephalography - event related neuralnetwork ras -Vidal : Contre , J. L (1992a , J. P., & ). An integrated Banquet . In Proceedings on categorization andprobabilitycontexteffects modelof temporal of potentials - 546 : Erlbaum . onneural nthoorks theinternational , NJ ). Hillsdale (pp. 541 joint conferena -Vidal order : Contre ras , timing , J. P., & , J. L (1992b ). Temporal , andprobabilitycontext Banquet .), Artificial . In I. Aleksander & :J . Taylor (Eds andcategorization effects on patternrecognition - 1890 : North-HoUand . 2 (pp. 1885 neural networks ). Amsterdam -Vidal contexts andprimingpreparation : Contre ras , J. L (1993a , J. P., & ). Learning temporal Banquet on neural . In Proceedings worldconference for patternrecognition modes of theinternational : Erlbaum . networks - 131 , NJ (pp. 126 ). HiUsdale of multimodal : Contre ras -Vidal , J. L (1993b , J. P., & ). Spectral timing and integration Banquet . 354 . Amsterdam : neural networks 350 In . . es . ( ) (Ed ), J Taylor Artificial systemic pp process . Elsevier -termmemoryin contextprocessing : Contre ras -Vidal , J. L (1994 , J. P., & ). Mediumandlong Banquet -hippo conference relations . In Proceedings of theinternational : A networkmodelof cortex campus . . 4, (pp. 647 : Erlbaum onneural networks - 654 , NJ , vol ). HilIsda1e -time : GUnther , , W. (1997b , J., Joulain , C, Revel ,A,& ). Space , J. P., Gaussier , P., Dreher Banquet In . G. Matthews : A neural basis of in franto - hippocampal orderandhierarchy system personality - 189 . : Elsevier onpersonality andemotion science ). Amsterdam (pp. 123 (Ed), Cognitive perspectives and ., & : Joulain C. (1997a ,R ). Learning , P., Muga , S., Arnaud , J. P., Gaussier , recognition Banquet : A neuralnetwork - hippocampal - spatialsequences of temporo system by a cortico generation : Neuralmodels on vision . In Proceedings model , action , recognition of mindand of the conference . ForAdaptiveSystems machine , MA: Center (p. 46). Boston es throughthe structure of : Grossberg , J. P., & ). Probingcognitiveprocess , S. (1987 Banquet -relate . Applied : An experimental andtheoretical event , analysis Optics duringlearning potentials - 4946 26 . , 4931 of task and stimulusprobabilityon : Lesevre , N. (1981 , J. P., Renault , B., & ). EHect Banquet - 214 . evoked . Biological , 13 , 203 Psychology potentials W. (1992 es andmotivation ., & : GUnther , attention , M. J , J.-P., Smith ). Top-downprocess Banquet in .), Motivation andg~ l direction in cognitivetasks . In D. S. Levine& : S. J. Leven(Eds , emotion . : Erlbaum neural networks - 207 , NJ ) . Hillsdale (pp. 169

Introduction to Neural Networks

: Renault , J. -P., Smith , M . J., & , B. (1990). Bottom-up versus top-down: An alternative to Banquet the automaticattendeddilemma ? Beh Rvioraland Brain Sciences , 13 , 233- 234. . S., & : Levine ,R , D. S. (1994). ModeDing the role of the frontal lobes in sequentialtask perBapi . l. Basicstructure and primacy eHeds fonnance . NeuralNetworks , 7, 1167- 1180. : RoDs of neuronsin the inferior temporal cortex in short, E. T. (1987). Responses Baylis, G. C., & term and serial recognition memory tasks . Experimental Brain Research , 65, 614- 1622. : Thompson , T. W., & , R. F. (1978). Neuronal plastidty in the limbic system during classical Berger . Brain Research conditioning of the rabbit nictitatin membraneresponse , 145 , 323- 346. Bradski : Grossberg , G., Carpenter , G., & , S. (1992). Working memory network for learning temporal order with application to three-dimensionalvisual object recognition. Neural Computation , 4, 270- 286. Brebion , J. (1994). UnpublishedPhiD. dissertation . , Universite ReneDescartes , Paris Brown , J. (1958). Some tests of the decay theory of immediate memory. Quarterly Journal of , 10 , 12- 21. Experimental Psychology : Grossberg BuDock , D., Fiala , J. C., & , S. (1994). A neural model of timed responselearning in the cerebellum . NeuralNetworks , 7, 1101- 1114. : O ' Keefe , N., Recce , M ., & , J. (1994). A model of hippocampalfunction. Neural Networks Burgess , 7, 1065- 1081. -HaD Burnod , Y. (1990). An adaptiveneuralnetworkthecerebral cortex . London: Prentice . Buzsaki , G. (1988). Polysynaptic long-term potentiation: A physiological role of the perforant . Brain Research , 455 , 192- 195. path- CAJ / CA1 pyramidal cell synapse Buzsaki , G. (1989). Two- stagemodel of memory trace formation: A role for " noisy" brain states . Neuroscience , 31, 551- 570. : Grossberg , G. A , & ). A massivelyparaDelarchitecturefor a self- organizing Carpenter , S. (1987a neural pattern recognition machine . Computer Vision , Graphics , and ImageProcessing , 37, 54- 115. : Grossberg , G. A ., & - organization of stable category recognition ). ART 2: Self Carpenter , S. (1987b codesfor analog input patterns . Applied Optics , 26, 4919- 4930. : Grossberg , G. A , & Carpenter , S. (1990). ART 3: Hierarchicalsearchusing chemicaltransmitters in self- organizingpattern recognition architectures . NeuralNetworks , 3, 129- 152. : Grossberg , G. A ., & Carpenter , S. (1993). Normal and amnesiclearning , recognition, and memory . Trends in Neuroscience by a neural model of cortico- hippocampalinteractions , 16 , 131- 137. : Reynolds , G. A , Grossberg , J. H. (1991). ARTMAP : Supervisedreal-time learning Carpenter , S., & and classificationof nonstationary data by a self organizing neural network. NeuralNetworks , 4, 565- 588. Cave : Squire , C. B., & , L R. (1992). Intact and long-lasting visual object priming in amnesic . Journalof Experimental , Learning , Memory, and Cognition , 18 , 509- 520. patients Psychology Cho, Y. H., Beracochea : Jaffard . (1992). Differential eHeds of ibotenate lesions of the , D., & , R CA1 subfield of the hippocampuson a delayed non-matching-to- place task as a function of preoperative , 20 (4), 261- 269. training in mice. Psychobiology Cohen : Squire . (1980). Preservedlearning and retention of pattern analysing skiDin , N. J., & ,L R amnesia : Dissociation of knowing how and knowing that. Science , 210, 207- 209. Corkin, S., & : Milner , B. (1968). Acquisition of motor skiD after bilateral medial temporal lobe exdsion. Neuropsychologia , 6, 225- 265.

115

radion ~ , and Cortico- HippocampalInt@ Banquetet al.: Memory, Amnesia

Dehaene : Nadal, J. P. (1987). Neural networks that learn temporal sequences , S., Otangeux , J. P., & . Proceedings AcRdmryof Sciences , 84, 2727- 2731. of the NRnonal by selection Dradtrnan, D. A ., & : Arbit , J. A (1966). Memory and the hippocampalcomplex. Archivesof Neurology , 15 , 52- 61. : Cohen Eidtenbaum , H., Otto , T., & , N. J. (1994). Two functional componentsof the hippocampal

. Behavioral andBminScima - 518 . , 17 , 449 memory system -Rakic Funahashi ., Bruce : Goldman ,S , C., & , P. W. (1989 ). Mnemonic codingof visualspace by 's dorsolateral neurons in the monkey by an oculomotor prefrontalcortex revealed delayed task . . , 61 , 1 19 of Neurophysiology response Journal Fuster in thecerebral cortu . Cambridge : MIT Press . , J. M. (1995 ). Memory , MA Fuster . Anatomy cortu andneuropsychology , J. M. (1997 , physiology ). Theprefrontal of thefrontal . 3rd ed . New York : Raven lobe Press . -termmemory Fuster : Alexander to short . Science , G. E. (1971 , J. M., & ). Neuronactivity related , 173 - 654. , 652 Fuster : Jervey . P. (1981 neurons andretainbehaviorally , J. M., & ,J ). Inferotemporal distinguish relevant features of visualstimuli . Scima , 212 , 55. Glanzer : Amitz mechanisms in free recall . Journal , M., & , A R. (1966 ). Two storage of Verbal - 360 andVerbal Behavior . , 5, 351 Learning -Rakic Goldman of cognition . Parallel distributed networks in primate , P. W. (1988 ). Topography - 156 cortex . Annual association Review . , 11 , 137 of Neuroscience . , Whitson : Lynm of long-term ,R , J., Larson , J., & , G. (1994 ). Non-Hebbian Granger properties enable <apadty encoding of temporal . Proceedings Academy of National high potentiation sequences - 10108 . , 91 , 10104 of Sciences : Singer in cat visual , C. M., Konig , A. K , & , W. (1989 ). Oscillatoryresponses Gray , P., Engel cortexexhibit intercolumnar which reflectglobalstimulus . Nature , synchronizations properties - 337 338 . , 334 anduniversal development ). Adaptivepatternclassification Grossberg , S. (1976a recording , I: Parallel - 134 andcodingof neural feature detectors . Biological . , 23 , 121 Cybernetics anduniversal ). Adaptivepatternclassification , expectation Grossberg , S. (1976b recoding , II: Feedback - 202 . Biological . , olfaction , andillusions , 23 , 187 Cybernetics : Self - organization and performance of ). A theory of humanmemory Grossberg , S. (1978 motor codes and . In R . Rosen & : F. SneO . in theoretical , , , (Eds ), sensory Progress maps plans - 374 : Academic Press . Reprinted in S . Grossberg .). , Vol. 5 (pp. 233 ). New York (Ed biology : Neural Studies , perception , development , cognition , andmotor of mindandbrain of learning principles control . Boston : Reidel Press . : MemO . (1992 timedreinforcement ,J ). A neuralnetworkmodelof adaptively Grossberg , S. & andhippocampal . Cognitive Brain Research ,1 , 3- 38. dynamics learning : Schultz of learningin the , E., Dorizzi , B. Bumod , Y., & , W. (1995 ). Neuralcorrelates Guigon - 147 cortexof themonkey : A predictive . Cerebral model Cortu . , 5, 135 prefrontal : Crandall , E., Squires , N. K., Wilson , C. L. Rohrbaugh , J. W., Babb , T. L., & , P. H., (1980 ). Halgren in the human formation andamygdala Endogenous generated potentials by infrequent hippocampal - 805 events . Scima . , 210 , 803 Horel of general . Behavioral andBrainSciences , J. A (1994 ). The localization , memoryfunctions 17 . , 482

Introduction to Neural Networks

: Keele Uum . Journalof CognitiveNeuroscience , S. W. (1989). Timing functions of the cerebe Ivry , R. B., & , 1, 134- 150. . New York: Holt . , W. (1890). ThePrinciples James of Psychology . , Jr. (1995). On the neural generatorsof the P300: Evidencefrom temporal lobectomy ,R Johnson . Electroencephalography and Clinical Neurophysiology , 44 (Suppl.) 110- 129. patients . , Jr., & : Donchin E. (1982). Sequentialexpectancesand dedsion making in a changing ,R Johnson environment: An electrophysiologicalapproach . Psychophysiology , 9, 183- 199. -hippocarnpalconnedions: A speculativeview of their function. , R. S. G. (1993). Entorhinal Jones Trends in Neurosaences 16 5864. , , Kim, J. J., & : Fanselow , M . S. (1992). Modality specific retrograde amnesiaof fear. Science , 256 , 675- 677. Levine : Parks , D. S., & , R. (1992). Frontal lesion effects on verbal Ruencyin a network model. In on neuralnetworks , vol. 2 (pp. 39- 44). of the international joint conference Proceedings La, Z. L., Wi Uiamson : Kauhnan , S. J., & , L. (1992). Human auditory primary and associationcortex . Brain Research have differing lifetimes for activation traces , 572 , 236- 241. : Squire . (1989). Autobiographical memory and amnesia . Psychobiology MacKinnon , D. F., & ,L R , 17 , 247- 256. ' McClelland, J. L., McNaughton . C. (1995). Why there are complementary : O Reilly, R , B. L , & : Insights from the success es and failures learning systemsin the hippocampusand the neocortex of the connectionistmodels of learning and memory. Psychological Review , 102 , 419- 457. : Horwitz L G. (1996). Changes Mcintosh, A. R., Grady, C. L , Haxby, J. V., Ungerleider , L G., & in limbic and prefrontal cortex functional interactions in a working memory test for facts . Cerebral Cortu, 6, 571- 584. Miller , E. K., & : Desimone , R. (1994). ParaDelneuronal mechanismsfor short-term memory. Science , 263 , 520- 522. Miller , E. K., Li, L , & : Desimone . (1993). Activity of neurons in anterior inferior temporal cortex ,R , 1460- 1478. , 13 during a short-term memory task. Journalof Neurosdena Milner , B. (1959). The memory defect in bilateral hippocampal lesions . PsychiatricResearch , 11 , 43- 52. Reports Milner , B. (1962). Les troubles de la memoire accompagnant les lesions hippocampiquesbilater' . In Physiologie : CNRS. ales de l hippocampe (pp. 257- 272). Paris Milner, B. (1966). Amnesiafollowing operation on the temporal lobes. In C. WM . Whitty & : O. L. Zangwill (Eds .), Amnesia . London: Butterworths. Milner , B., Corkin : Teuber , S., & , H.-L. (1968). Further analysisof the hippocampalamnesiasyndrome : 14-year follow -up study of H. M . Neuropsychologia , 6, 215- 234. : Chang Miyashita, Y., & , H. S. (1988). Neuronal correlate of pictorial short-term memory in the cortex. Nature , 68- 70. , 331 primate temporal N Utinen , R. (1982). Processingnegativity : An evoked-potential reRedion of selectiveattention. Bulletin , 92, 605- 640. Psychological Niki , H. (1974 ). Differential activity of prefrontal units during right and left delayed response trials. Brain Research , 70 , 346- 349. ' as a cognitivemap . Oxford , U.K.: Oxford University O Keefe : Nadel, L (1978). The hippo , J., & campus Press .

117

, and Cortico-HippocampalInteractions Banquetet al.: Memory, Amnesia

Olton, D. S., Becker : Handelmann , J. T., & and memory , G. E. (1979 . ). Hippo , space campus - 365 andBrain Behavioral Science . , 2, 313 Owen . G., Sakakian , A M., Morris : Robbins ,R , R., Polkey , C. E. & , T. W. (1996 ). Doubledissociations of memory andexecutive functions in workingmemorytasks followingfrontallobe - 1615 excision lobeexcision or amygdalo - hippocampedomy , temporal in man . Brain . , 119 , 1597 Parks . W., Levine ,R , D. S., Long , D. J., Dalton , I. E., Weingarther , H., Fedio , P., , D. L., Crockett : Becker Coburn , K. L., Siler , G., Matthews , J. R., & distributed , R. E. (1992 ). Parallel processing andneuropsychology : A neuralnetworkmodelof Wisconsin CardSortingandverbalfluency . - 233 Review . , 3, 213 Neuropsychology Paulesu : Frackowiak . S. J. (1993 , E., Frith , C. M., & correlates of the verbalcomponent ,R ). The neural - 345 of workingmemory . Nature . , 362 , 342 -termretention Peterson : Peterson , L. R., & of individual , M. J. (1959 verbalitems . Journal ). Short 58 193 198 . , , of Erptrimental Psychology Petrides : Meyer , M., Alivisatos , B., Evans of humanmid-dorsal , A., & , E. (1993 ). Dissociation from postero -dorsolateral frontalcortexin memory . Academy of theNational Proceedings processing - 877 . , 90 , 873 of Scienas - 787 Reiss : Taylor . Neural , M., & Networks , J. (1991 . ). Storingtemporal , 4, 773 sequences Rolls . T. (1989 of neural networks in the hippo ,E andneocortex in memory ). Fundions . campus In J . H. Byrne& : W. O. Berry(Eds .), Neural models : Theoretical andempirical OR L 'hes of piR Sticity appr : Academic . . Press (pp. 240- 265 ). NewYork RoDs . T., Baylis : Nalwa ,E , G. C., Hasselmo , M. E., & on theface , V. (1989 ). Theeffectof learning selective of neurons in the cortex in the superiortemporalsulcus in the monkey . responses - 164 BrainRt StRrch . , 76 , 153 Erptrlmental -related Ruchkin ., Johnson . , Jr., Canoune : Ritter , D. S ,R , H., & , W. (1991 ). Event potentials during - 487 arithmetic andmental rotation . Electroencephalography andClinical . , 79 , 473 Neurophysiology Ruchkin . , Jr., Grahnan ., Canoune : Ritter , D. S., Johnson ,R ,J and , H., & , W. (1992 ). Distinctions similarities es : An event related . Cognitive amongworking memoryprocess potentialsstudy BrainRtse Rrch - 66. ,1 , 53 Schader : Ochsner , D. L , Chu , C.-Y., & : A selective , K N. (1993 review . ). Implicit memory Ann UR IReview - 182 . , 16 , 159 of Neuroscience Schader : Graf , D. L , & , P. (1986 : Perspectives ). Preserved from learningin amnesic patients - 743 research on directpriming . Journal andErptrlmental . , 6, 727 of Clinical Neuropsychology Scoville : Milner afterbilateral , W. B., & , B. (1957 . lesions ). Lossof recent memory hippocampal - 21. , Neurosurgery , andNeuropsychiatry , 20 , 11 Journal of Neurology Shiffrin . M., & : Schneider Dedand automatic ,R , W. (1977 . II: ). Contro information processing - 190 anda general . Psychological Review . , 84 , 127 Perceptual learning , automatic attending theory Shimamura : Squire . (1987 , A P., & ,L R andsource ). A neuropsychological studyof fad memory - 473 amnesia . Journal : Learning andCognition . , Memory , 13 of Erptrlmental , 464 Psychology Sidman : Mohr, J. P. (1968 , M., Stoddard additional , LT ., & observations of ). Some quantitative - 254 immediate in a patientwith bilateral . Neuropsychologi . , 6, 245 memory R hippocampailesion , W. (1991 - organization ). Neuronaladivity as a shapingfador in the self of neuron Singer . In G. ACarpenterS . Grossberg assemblies .), PaHem neural (Eds recognition by selforganizing networks . Cambridge . , MA: MIT Press for everything ? Temporal Solomon , P. R. (1980 ). A timeanda place viewsof hippo processing - 261 function with special reference to attention . Physiological . , 8, 254 campal Psychology

118

Introduction to Neural Networks

. (1992 ,L R . A synthesis ). Memoryandthe hippo from findingsin rats Squire , monkeys campus - 231 andhumans . Psychological Review . , 99 , 195 . , Cohen : Nadel , L. R , N. J., & , L (1984 ). The medial temporallobe and memory Squire consolidation : A new hypothesis . In H. Weingartner & : E. Parker .), Memoryconsolidation (Eds - 210 : Erlbaum . (pp. 185 , NJ ). Hilisdale . , Haist : Shimamura , L R , F., & , P. (1989 : Quantitative ). The neurologyof memory Squire assessment of retrograde amnesia in two groupsof amnesic . Journal ~ dtnct , 9, of Neur patients 828 - 839 . : Musen , L R., Knowlton , B., & and organization . G. (1993 of memory ). The structure . Squire - 495 Ann U R I Review . , 44 , 453 of Psychology . , Ojemann ,L R , J. G., Miezin , F. M., Petersen : Raidtle , S. E., Vdeen , T. 0 ., & , M. E. (1992 Squire ). Activationof the hippo in nonnalhumans : A functional anatomical . campus study of memory - 1841 Ct S . , 89 , 1837 Proceedings of theNational Academy of Scitn . C., Wickens ,K : Donchin , C., Squires , N. C., & . E., (1976 Squires ). Theeffectof stimulus sequence -related - 1146 on thewaveform of cortical event . Science 193 . , , 1142 potential -termpotentiation Staubli : Lynch , U., & , G. (1987 ). Stable elicited hippocampal long by thetapattern - 234 . Brain stimulation Research . , 435 , 227 : Rudy Sutherland , R. W., & , J. W. (1989 association : The role of the hippo ). Configural theory formation in learning . Psychobiology , andamnesia - 144 . , 17 , 129 campal , memory . New York : Oxford UniversityPress ). Elmrents . Tulving , E. (1983 of episodic memory : McCarthy , E. K., & of retrograde , R. A. (1988 amnesia . Brain and ). Thefractionation Warrington 7 184 200 . , , Cognition : Weiskrantz . E. K., & of the prior learning , L (1978 effectin amne ). Further Warrington analysis - 177 sicpatients . Neur . , 16 , 169 O P S .vchologia andconsolidation , W. A (1979 : A theoretical ). Chunking of semantic networks Wickelgren synthesis -R versus , configuringS , and cognitive learning , nonnalforgetting , the amnesic syndrome the hippocampal arousal . Psychological Review , 86 , 44- 60. system Wilson : McNaught , M. A , & of the hippocampal on . B. L (1993 ensemble codefor ). Dynamics . Science 261 1055 1058 . , , space Winocur amnesia in rats with dorsalhippocampal , G. (1990 ). Anterograde and dorsomedial - 154 thalamic lesions . Behavioral BrainResearch . , 38 , 145 -MorganS., & Zola : Squire formation : Evidence , L (1990 for a limited ). Theprimate hippocampal rolein memory . Science - 290 . , 250 , 288 storage -MorganS., Squire Zola L, & : Amaral amnesia and the medialtemporal , D. G. (1986 ). Human : Enduring lesionlimitedto the fieldCA1 of the region memory impairment followinga bilateral . Journal - 2967 . , 6, 2950 of Neuroscience hippo campus

119

, and Banquetet aI.: Memory, Amnesia

Interactions

II

Behavioral

States

A Computational Model of Alcohol Dependence: Simulation of Genetic Differences in Alcohol Preference and of
Therapeutic Strategies

Raymond L. Ownby

on alcohol use disordersduring the past severalyears has provided Research of alcohol important information on the possible neurobiological mechanisms exposure that chronic ethanol for . have shown Studies , , example dependence affectsthe G proteins important in cellular secondmessenger systems : , 1992) and may also affect membrane calcium channels (Dolin & ( Nestler Patch Little, 1989; Dolin , Little , Hudspith, et al., 1987; Whittington , Dolin , , shows that acute and chronic ethanol exposureaffect et al., 1991). Research neural transmission , especiallythe dopaminergic systemsof the mesolimbic : Imperato Chamess 1992 , Acquas , , 1988a , b; Di Chiara , ; Di Chiara & system ( Cador & : LeMoal : Di Chiara i , 1992; Imperato & & : Carbon , , , , 1986; Stinus : Lin1992;) and serotonergic systems in other areas (George , Wozniak , & noiia, 1992; McBride, Murphy , Lumeng , Pihl, , et al., 1992; LeMarquand . of alcohol 1992 Strains & : Sobell Sellers 1994 & : Benkelfat , , Higgins, ) ; , and differ in rats for , , expression example preferring and nonpreferring ens ( NAcc ) and regulation of dopaminergic systems in the nucleus accumb the ventral tegmental area ( VTA ) and in the activity of their serotonergic , et al., 1987). Genetic differencesin the , McBride, Lumeng systems( Murphy ' rats dopamine system to alcohol consumption have been responseof these . demonstrated Similar differencesin serotonergic function are implicated in the susceptibility to developing at least one type of alcoholism (George et al., 1992). Still other studieshave suggestedthat brain systemsthat use y-aminobutyric ' acid (GABA ) as a neurotransmitter are important mediators of alcohol s ' . To make matters even more complex, at least part of GABA s reinforcing effects effectsare causedby its inhibitory effectson dopaminergicpathways : Wise, 1993; Terenius in the VTA (Devine, Leone , 1994). By inhibiting ,& these pathways , which are in turn inhibitory to the NAcc, alcohol stimulates increaseddopaminergic activity in the NAcc. Other lines of research show that glutamatergic neurotransmission , and especially the N-methylin alcohol' s reinforcing effects is involved te ( NMDA ) receptor o- asparta , , Fitzgerald , , 1992; Trevisan , et al., 1994; Chamess , Heinrichs (Koob, Rassnick Brose , et al., 1994). Limited data suggestthat homotaurine, a GABA agonist and a glutamate antagonist , may be effective in reducing alcohol intake in

rats (Koob, 1992a , b), but extensive studies have not been done in humans (Lhuintre, Daoust, Moore, et al., 1985). Neurobiological research thus implicates a number of neurotransmitter systems in the development and maintenanceof alcohol dependence . This researchhas led to new treatment strategiesfor the disorder. Studies show that opiate antagonist medicationssuch as naltrexone reduce alcohol intake in animals and decreasesubjective desire to drink and voluntary alcohol ' ingestion in humans(Volpicelli , Alterman, Harashida , et al., 1992; O Malley , , Change Jaffe , et al., 1992). Similar studies with serotonergic medications suchas the serotonin uptake blocker fIuoxetine have shown that thesemedications may also reduce the subjective desire to drink and lower alcohol intake in humans(LeMarquandet al., 1994). Koob and his associates : Bloom, 1988; Koob et al., (Koob, 1992a , b; Koob & 1994) have researched the neural circuitry that probably underlies alcohol' s that alcohol and other abusedsubstances reinforcing properties. Koob emphasizes - a neural circuit including act on critical brain reinforcement systems most importantly the NAcc and VT A but other structures also (figure 5.1). Acute alcohol use results in increaseddopaminereleasein the NAcc, an event that may be a neurophysiological correlate of behavioral reinforcement . Chronic alcohol use induceschangesin this system so that when alcohol is not present it is activated . This may be the neurophysiological counterpart of the increaseddesire to consumealcohol that is subjectively " " experiencedas craving.

Figure5.1 Elementsof the neural circuit underlying alcohol' s reinforcing eHeds.

BehavioralStates

: Beresford Modell and his colleagues(Modell , Mountz , & , 1990) have also reviewed this circuitry . They emphasize , however, the similarity of obsessive -compulsive disorder (OCD ; Modell , thoughts about drinking and obsessive Glaser , Mountz , et al., 1992b 1993). Modell ; Modell , Glaser , Cyr, et al., 1992a and his colleaguesargue that the samefrontal cortex- dorsomedial thalamic circuit implicated in OCD symptoms may also mediate obsessivethoughts . They suggest that activity in about alcohol in alcohol-dependent persons this corticothalamic loop is regulated by an inhibitory circuit that includes the NAcc and ventral pallidum. Inhibitory serotonergic inputs from the dorsal raphe nuclei also modulate activity in theseloops via input to the NAcc. Ethanol intake acutely increases dopaminergicactivity in the NAcc, reducing inhibition in the NAcc ventral pallidum circuit. The result of this combination of excitation and inhibition is increasedactivity in the frontothalamic . Increasedactivity in this frontothalamic circuit circuit implicated in obsessions results in subjectively experienced obsessive thoughts about alcohol. . The behavioral outcome of activation in this circuit is alcohol consumption These analyseshave considerableheuristic value and are solidly based in . They are the most comprehensive neuroanatomicaland neurophysiological research involved in alcohol current formulations of the neural mechanisms about the neurobiology much current and , knowledge integrate dependence . As proposed theories of alcoholism ; , though, they are static and conceptual A in . suitable are not amenable to , dynamic they testing although interesting vitro model of the neural elements involved in alcohol dependencewould allow dynamic testing of hypotheses about etiology and treatment , but at . Computational modeling of anatomicalneural presentis technically unfeasible networks, however, provides a way to examine relations among neural structuresimplicated in symptoms of alcohol dependence , such as changesin the desire to drink or thoughts about drinking . Such models can represent underlying physical systemsin ways that allow experimentsthat are otherwise the effectsof new medications difficult or impossible(such as assessing on neurotransmitter systems ). For further discussionof applications of comfulness , seechapter 1 in this book. putational modeling and their use

METHOD
Model Development
The anatomicalmodels developed by Koob et al. (Koob, 1992a , b; Koob & Bloom, 1988; Koob et al., 1994) and Modellet al. (Modellet al., 1990) were the basis for the network createdin this study (seefigures 5.1 and 5.2). The of the biological circuitry simulatedare (a) the changes central characteristics in dopaminergic responsivity in the NAcc and VT A system accompanying alcohol intake that are under genetic control in certain strains of rats; and (b) the dorsomedial thalamus and orbitofrontal cortex system , believed to be al. Mountz et Modell Curtis 1989 in both OCD overactive , , , , ) and alcohol (

Ownby : A Computational Model of Alcohol Dependence

LIMBIC CORTEX

NUCLEUS VENTRAL DORSOMEDIAL PALLIDUM ACCUMBENS THALAM US

OLFACTORY CORTEX

PONTINE NUCLEUS

AMYGDALA -A- = INDICATES

DORSAL RAPHE

I A

-RECURRENT SELF LINKS

FRONTAL VENTRAL TEGMENTAL CORTEX AREA

Figure5.2 Structureof the computationalnetwork simulating the circuit. dependence. This second circuit is believed to underlie the subjectively experienced intense desire to consume alcohol , similar to the intense obsessions and irresistible compulsions experienced by patients with OCD ( Modell et al., 1990 ). A target symptom of alcohol dependence was chosen to simulate the desire to drink an alcoholic beverage . It was chosen for several reasons. First , it can be argued that the desire to drink alcohol can be experienced at high intensities by persons who are not dependent on or abusers of alcohol (e.g ., in the context of significant thirst after working outside on a hot day , when someone might anticipate that a cold beer would be very refreshing ). It is assumed that a person not dependent on alcohol would have a desire to drink that is qualitatively similar to that of persons who are. When this desire becomes abnormally intense in relation to environmental cues and the likely social and medical consequences of drinking , it is pathological . The same network may thus be trained to simulate both a normal and an abnormal desire to drink alcohol . Second, although the relation between experienced desire to drink and actual drinking behavior is inexact , it can be argued that it exists (Bauer, 1992 ). Simulation of level of desire for alcohol is thus a meaningful goal . Third , at a more integrative level , the numerical value arbitrarily assigned to represent levels of desire can be understood to reflect the level of activation in the network produced by inputs and other ' experimental manipulations . Although it is not certain that the network s output to the dorsomedial thalamus or pedunculopontine nuclei corresponds

126

BehavioralStates

Table 5.1 Training set illustrating inputs and outputs BiasUnits


.1.1.1.1 .2.2.2.2 .3.3.3.3 .4.4.4.4 .5.5.5.5 .6.6 .6 .6 .7.7.7.7 .8.8.8.8 .9.9.9.9 1111

Limbic Cortex
0001 0010 00 11 0 100 0101 0110 0111 10 0 0 1001 1010

Olfactory Cortex
0001 0010 0 0 11 0 10 0 0101 0110 0111 1000 1001 1010

Amygdala
0001 0010 00 1 1 0 100 0101 0110 0111 1000 1001 1010

Dorsal Raphe
- .1 - .1- .1- .1 - .2 - .2 - .2 - .2 - .3 - .3 - .3 - .3 - .4 - .4 - .4 - .4 - .5 - .5 - .5 - .5 - .6 - .6 - .6 - .6 - .7 - .7 - .7- .7 - .8 - .8 - .8 - .8 - .9 - .9 - .9 - .9 - 1- 1- 1- 1

Outputs
0.10 .1 0.20 .2 0.3 0.3 0.4 0.4 0.50 .5 0.60 .6 0.70 .7 0.8 0.8 0.90 .9 1.01 .0

to a subjective desire to drink, it is likely that it is related to the behavioral activation resulting from presentation of alcohol -related cues (Dolinsky , : Stitzer , 1989 ). Reported Morse , Kaplan , et al., 1987 ; Turkkan , McCaul , & desire to drink alcohol can thus be understood to be the experienced counterpart of activation in the frontothalamic circuit . Activation can be read out ' at the dorsomedial thalamus and the pedunculopontine nucleus. A person s report of this readout , however , is affected by many other variables , including memory , environmental cues, and attentional focus. The network can thus be broadly conceived as processing sensory and motivational inputs from the olfactory and limbic cortices and the amygdala , under the influence of serotonergic and dopaminergic inputs . The output of the network represents different levels of activation of the circuit , corresponding to subjectively experienced obsessive thoughts about drinking alcohol , often " " referred to as craving . To develop a model of the pathological Overall Strategy for the Model first alcohol it was need to drink , necessary to develop a neural network that could simulate normal desire. The network would associate low levels of sensory and motivational cues with low levels of desire to drink, and higher levels of cues with higher levels of desire. The data used initially to train the network were arranged so that inputs from limbic and olfactory cortices and the amygdala conveyed gradual increases in value , conceptually associated with increased intensity of cues. Inputs from the dorsal raphe were increasingly negative , consistent with the inhibitory function of serotonergic projections from it to the NAcc . Inputs from bias units to the NAcc were increasingly positive , consistent with increases in dopaminergic activation associated with more intense sensory and motivational cues (table 5.1). Inputs from these four systems to the NAcc are modeled . These systems were chosen because of their importance in alcohol -reinforcing mechanisms, : Bloom , 1988 ; as discussed by Koob and colleagues (Koob , 1992a, b ; Koob &

127

: A Computational Modelof AlcoholDependence Ownby

Koob et al., 1994) and Modellet al. (1990). The actual network was constructed with four input units for each system , for a total of sixteen input units, with an additional four bias units. The next layer of this four-layer network comprised six units, corresponding to the anatomicalNAcc. These six units are connectedto six context units; this interactive system simulates the reciprocal anatomical interconnectionsof the VT A and the NAcc. The next layer of six units represents the ventral pallidum, again according to the anatomicalhypothesesof Koob and Bloom (1988). It in turn has six context units connectedto it , simulating the reciprocalanatomicalconnectionsto the frontal cortex area thought to be important in alcohol' s reinforcing properties . Note that this frontal cortex area is not believed to be the sameorbito frontal areathat combineswith the dorsomedialthalamusin the overactive circuit underlying obsessionsin OCD or obsessivethoughts about alcohol in alcohol dependence (Modellet al., 1990; Modellet al., 1989). The output are modeled with one unit (eachunit is represented systems by a black dot in 5.2 . These ) figure represent the dorsomedial thalamus and the pedunculo , both important targets because of their relevanceto behavioral pontine nucleus activation. The network thus has a partially recurrent , or Elman architecture (Hertz, & : Palmer . 1991 , ) Although this architectureis typically used to learn Krogh, tasks that require learning over time, it is used to here to create a network structure that is the analog of the underlying anatomicalnetwork. A sequence of inputs was developedto train the network and is illustrated in table 5.1. Low levels of cue intensity are representedinitially by " 0 0 0 1" (binary one) with increasinglevels of cue intensity ranging through " 0101" " " (binary 5) to 1 0 1 1 (binary ten). Dorsal raphe inputs decrease in smaller increments (becoming more negative) while bias unit inputs increase in the same increments (becoming more positive). These ten inputs reflect a of statesof activation of the neural system important in mediating sequence desire for alcohol. The first severallines of table 5.1 show what occurswhen motivational and sensorycuesare weak- the level of desire to drink is corresponding low . The next severallines illustrate changesin these parameters associatedwith increasing intensity of cues , and the desire to drink increases . The last lines show what might happen when cues are more . The personmay experiencea desireto drink that overwhelms inhibitory intense influences .
A network with Training Procedure and Post - Training Assessment twenty input units was created, with four units each for the four systems and four bias units (see figure 5.2 ). Six units with six corresponding and reciprocally interconnected context units were created in the two middle , or hidden , layers ; these were analogous to the processing elements in the NAcc and VT A and the ventral pallid um and frontal cortex circuits . One output unit was created for each output of the system . The network was trained using a backpropagation algorithm (see Hertz et al., 1991, and Parks, Levine ,

128

Behavioral States

& : Long, chapter 1 in this book). The model was then trained to simulate various levels of desire to drink, using the backpropagationalgorithm described in detail elsewhere (Rumelhart : Williams, 1986; see , Hinton , & chapter 1). Briefly, in a backpropagationnetwork, the model is presentedwith a set of matchedpairs of inputs and their correspondingdesired outputs. Training of the model begins when inputs are propagated forward through the several levels of neural elements . Firing of individual units is determinedby the synaptic ' weights between the elements and the unit s firing threshold. In this case , weights are analogsof the strength of the connectionbetween two neurons , and thresholdsare analogsof the critical voltage required for a neuron to generatean action potential. Weights are initially assignedrandomly. Inputs are then propagated forward to the output where some value . This actual output of the network is compared with desired output results (already specified ), and the error is calculatedas the differencebetweenactual and desired outputs. This error is then propagated back through the network , resulting in adjustmentsto the synaptic weights. The cycle is repeated for each pair of inputs and outputs until the network performs at a desired level of accuracy , usually at less than 1 percent mean squarederror, or MSE Rumelhart et at., 1986 , for a more detailed explanation (see ). As initial seedingsof synaptic weights are random and backpropagation learning employs a minimization of error algorithm, it is possible that any one network could arrive at a solution based on a spurious local minimum. To control for this problem, five networks were constructed and tested in eachexperimentalcondition describedbelow. With the small number of outputs in experimental conditions (ten for each ), it was possible that the the use of statistics would be violated in assumptionsunderlying parametric tests of differencesbetween network outputs. In these casesnonparametric tests were used to assessdifferences in network outputs, with a nominal : Castellan , 1987). In thesecases , two -tailed tests alpha value of 0.05 (Siegel & were employed throughout. In one experimentalcondition in which the rate of learning in two types of networks was compared over multiple trials, . This analysisof variance(ANOV A ) was used parametricrepeatedmeasures of the large number of data points used (four per appearedjustified because network times fifty networks) and statisticalevidencethat parametricassumptions were met ( Winer , 1971; discussedfurther below). In order to assess whether results were caused by local minima or other factors associated with one network, a total of five networks was constructedand trained. Eachof the five networks was easily trained to an MSE level of lessthan 1 ' . After training to assess the network s learning, each was tested by percent presenting it with the sametraining inputs but without the desired outputs. Trained outputs were examinedto determine whether the network operated as intended. Outputs of the five networks after training were not significantly different Wilcoxon Matched-Pairs Signed Ranks tests ; all probabilities > 0.10). To (

129

Ownby : A

I Model of Alcohol Dependence

: This table presentsoutputs from eadt network after training. Eachnet was presentedwith Note the sameset of inputs (the training set in table 5.1, but without the output) and its output was recordedto assess whether eachnetwork' s output was substantially the same .

further assess the extent to which all five networks behaved the same , new test inputs were constructed and presented to the networks. These were constructed on the rationale of holding the input of four of the systems constant and varying the fifth. This was an important test, since several of the planned experimental manipulations of the networks would assessthe ' changesin the networks outputs causedby changesin inputs. The outputs of eachof the five networks for eachof five conditions were comparedusing -Sum tests. They did not differ significantly Wilcoxon Matched-Pairs Rank P .10 > table 5.2 . All five networks were then tested in all subsequent ( ) ( ) conditions. Unless otherwise noted, the outputs did not significantly differ among the networks acrossconditions. The study presented in this chapter used neural network software developedat the University of Stuttgart Institute for Paralleland Distributed : Vogt , 1996). It was run on a High PerformanceSystems(Zell, Mamier, & Pentium computer using the Linux operating system ( Volkerding : , Richard ,& , 1996). Johnson
Test Conditions

Experimentalmanipulations to simulate alcohol-related desire to drink were of three kinds: (1) changesin inputs to the bias units to the NAcc to simulate changes in dopaminergic activity in it to simulate normal and abnormal levels of desire for alcohol- n9rmal and abnormal levels of desire to drink; (2) changes in the self-recurrent weights in the context units to simulate inherently higher levels of dopaminergic activity in those systems to simulate increasedgenetic vulnerability to developing alcohol dependence ; and (3) changesin inputs to the networks to simulate changesin the state of the system each input represents to simulate the effects of therapeutic interventions on the network.

130

BehavioralStates

Changes in Bias Units During training, the four bias units were each set to increaseby 0.1 unit to simulate the increasein dopaminergicactivation of the NAcc with changesin cue intensity. To investigate whether increased dopaminergicactivity in the NAcc would affect the level of desire to drink. inputs were changed to 1.0 for all cue levels. To simulate decreaseddopaminergic activity resulting from blocking dopaminergic inputs by opiate , in another condition inputs were changedto - 1.0. antagonist medications Changes in Self- Recurrent Weights In the architectureof the network as constructed , eachof the six context units was reciprocally interconnectedto simulatethe NAcc-VT A circuit. Eachcontext unit in turn has a self-recurrent connection , so that during training part of its own output is fed back to it . These self-recurrent weights are usually set to zero (Zell et al., 1995). To simulate increaseddopaminergic activity in this circuit, these weights were increasedto O .8- in this way, making the circuit inherently more active. In the normal condition, the network simulated alcohol-nonpreferring rat strains , while with increased dopaminergic activity it simulated alcoholpreferring rats. The rates of learning in two groups of twenty -five networks, one with normal dopaminergicactivation and one with increasedactivation, were comparedat four points via repeatedmeasures ANDY A and post hoc I-tests . as well under conditions Therapeutic Conditions The network was assessed interventions in alcohol . Opioid simulating therapeutic dependence medications have been in of alcohol effective treahnent antagonist dependence , presumably by antagonizing .u and K opiate receptors in the NAcc , 1988b (Oi Chiara & Imperato ). This antagonism results in decreaseddopaminergic activity in the NAcc. Although simulating changesin the inherent activity of the network via changesin the self-recurrent weights as described above was possible , it was not possible nor would it be realistic to manipulate these weights after learning had occurred in the network. The effect of decreaseddopaminergic activity in the NAcc was therefore simulated by decreased inputs to the NAcc from the bias units. Serotonin reuptakeinhibi. The possible effect on tors have also been used to treat alcohol dependence the serotonergic system of the use of these medications was simulated by . changing the serotonergicinput to the system from the dorsal raphe in treahnents are effective alcohol , Finally psychosocial treating dependence , et al., 1989). The (Marlatt & Gordon, 1985; Monti , Abrams, Kadden neurophysiological mechanismby which psychosocialtherapieswork is not , and different psychosocial therapies may operate on different readily apparent . Coping skills therapy might increase neurophysiological substrates frontal cortex control over drinking behavior, for example , while the understanding of the determinantsof drinking behavior resulting from participation in Alcoholics Anonymous might decrease the intensity of motivational . In to the this the , system study inputs impact of psychosocialtherapy was

Ownby : A ComputationalModel of Alcohol Dependence

simulatedby decreasinglimbic system inputs to the system , consistent with the hypothesis that psychosocialtherapy decreases the intensity of motivational inputs to the NAcc. Although this approach to simulating psycho' , it should be noted that Baxter and his therapy s effects is speculative Baxter Schwartz , , ( , et al., 1992) showed that both pharcolleagues Bergman and behavior for OCD macological changedmetabolic rates in ana therapy tomical structuresthat are parts of the circuit hypothesizedhere to be central to alcohol dependence .
Experiment " One : Normal " Desire to Drink

Networks were constructedand trained as described . To assess whether each of the five networks behavedsimilarly after being trained, the outputs of each network at the end of training were statistically compared . The behavior of each network in other conditions was then tested with new input patterns. These new inputs consistedof a set of five inputs from each system (limbic and olfactory cortices , amygdala , dorsal raphe , and dopaminergicbias units). In each group, one input was set at the level corresponding to the highest level of desire to drink, with all others at the minimum. In this way, the effects of inputs from individual anatomicalsystemson each network could be assessed . This was done becauselater experimental conditions would these manipulate inputs and it was therefore essentialto determine whether the experimentalinterventions affectedeachnetwork in the sameway. Each of the resulting outputs was tested for equivalence . If outputs from each of five networks were not significantly different acrossmultiple conditions, solutions were probably not basedon local minima.

" " ExperimentTwo: Abnormal Desire to Drink


The ability of the network to simulate very high levels of desire to drink in response to minimal alcohol - related stimuli (as occurs in alcohol -dependent . persons; Rohsenow , Niaura , Childress , et al., 1990 1991 ) was then assessed It has been hypothesized that alcohol related stimuli cause an increase in dopaminergic activity in the NAcc with a resultant loss of inhibition to the frontocortical - thalamic circuit involved in the subjective desire to drink ( Modellet al., 1990 ). Inputs to the bias units , used to simulate levels of dop aminergic activity in the NAcc , were increased to simulate this experimental " " condition . At the same time , the same inputs (simulating normal environmental stimuli ) were presented at other network inputs . Experiment Three : Genetic Vulnerability

Several neurochemical differences have been demonstrated in strains of alcohol-preferring and -nonpreferring rats. The most salient of these is the increasedresponseof their NAcc-VT A dopaminergic system to the admin-

BehavioralStates

istration of alcohol (Charness , 1992 ; Welss , Lorang, Bloom, et al., 1993). The architectureof the current network, with several layers of hidden units and corresponding groups of context units (see figure 5.2), allowed for simulation of this increasedpattern of reactivity . Eachof the context units (seeA in figure 5.2) has a self-recurrent link that is usually constrainedto be zero. By increasing this value to a positive nonzero value, the inherent activity of the circuit is increased , similar to the activity in this circuit inalcohol preferring strains of rats. It was hypothesizedthat with increasedactivity in this circuit, the network would more rapidly and completely acquiremultiple levels of desire to drink. To assessrate and completenessof learning, tIle ' networks level of error (MSE) was assessed at four learning points: after 500, 1000 , 1500, and 2000 learning trials. To assess adequatelywhether differences were not due to random differencesand to allow a larger data set for parametric analysis in , twenty -five networks were trained and assessed eachof the two conditions (low and high dopaminergicreactivity ).

ExperimentFour: TherapeuticModalities
Animal and human studies have shown that serotonin reuptake blocking and opiate antagonist medications can reduce alcohol intake , and in humans, reduce reported desire to drink (Anton , 1995 ). Psychosocial therapies , such as Alcoholics Anonymous , relapse prevention (Marlatt & : Gordon , 1985 ), and coping skills therapy (Monti et al., 1989 ) are effective in treating alcohol dependence. The effects of simulations of these treatments on levels of desire to drink were therefore studied . " " In all cases , alcohol -preferring networks were used for tests of therapeutic modalities . Abnormal desire to drink was simulated as outlined in experiment two (above ), and inputs were varied to simulate the impact of therapeutic interventions . To simulate the use of serotonin reuptake blockers , dorsal raphe inputs were increased by a factor of 10. The effects of opiate antagonist medications (decreases in dopaminergic activity in NAcc in response to stimuli ) were simulated by decreasing inputs &om the dop aminergic bias units . Psychosocial therapies were hypothesized to reduce motivation to drink; this decrease in motivation was simulated by reductions in limbic system inputs to the system .

RFSUL TS " " ExperimentOn ~: Normal Levelsof Desire to Drink


The hypothesis that the network could learn to simulate " normal" desire to drink alcohol, and even high levels of desireassociated with multiple internal and external cues and motivational states was tested first. The output of " the network in its untrained state was compared with its output after 2000 training trials with a resulting MSE less than 1 percent . The network

133

: A Computational Mode ! of AlcoholDependence Ownby

simulated " normal" levels of desire for alcohol: the two sets of output differed significantly (z = - 2.80, P < .01). " " Experiment Two : Abnormal Levels of Desire to Drink
The hypothesis was also tested that abnonnally high levels of desire to drink , in response to minimal cues, could be simulated by the network . With dopaminergic input from bias units to the NAcc increased, the same stimuli representing internal and external cues and motivational states were again ' presented to the network . The network s output was significantly increased (i.e., indicated higher levels of desire to drink ) in this condition (z = 2.70; P < .01 ). Experiment Three : Genetic V ulnerabilit }'

As outlined above , increaseddopaminergicactivity in the NAcc-VT A circuit was simulatedby changesin self-recurrent weights (seefigure 5.2), and rates of learning of " nonpreferring " and " preferring" networks were compared . five networks in each condition were assessed at four time Twenty points. ' Repeated measuresANDV A showed a main effect of time (Hotelling s T2 = 20.21; F = 227.41; df = 4.45; P < .001), as expected with effective training of both networks (figure 5.3). There was a significant interaction of group with time (Hotelling ' s T2 = 1.80; F = 20.23; df = 4, 45; P < .001), suggesting that each group had a different learning rate. The graph of the MSE for eachcondition (seefigure 5.3) illustrates the slower learning rate for the nonprefer ring network. Post hoc tests of these mean differences show that although each group of networks started at approximately the same level of error (comparison of baseline MSEs before training each network; t = - 0.25, df = 48; P > .10), their MSEs differed at each subsequentpoint (t-value at 500 trials = 9.19, df = 48; P < .001; I-value at 1000 trials = 7.01, df = 48; P < .001; I-value at 1500 trials = 4.69, df = 48; P < .001; I-value at 2000 trials = 3.61, df = 48; P < .001; all values significant after Bonferroni correction for multiple comparisons ). The significant differenceobtained after 2000 learning trials suggeststhat the preferring networks not only learned to desire alcohol more readily but also more completely, as might occur in a . person genetically vulnerable to the development of alcohol dependence

ExperimentFour: TherapeuticModalities
In each case , simulation of therapeutic modalities resulted in significant decreases in level of desire to drink. Opiate antagonist treatment reduced simulated desire to drink (all tests were Wilcoxon Matched-Pairs Signed Rankstests ; here , z = - 2.80; P < .01), as did treatment with serotonin reuptake blockers (z = - 2.80; P < .01). Psychosocialtreatment alone reducedthe desire to drink significantly (z = - 2.70; P < .01), but this decreasewas

134

BehavioralStates

0.250 0.005
~

0 . 2401 n.s. I0 . 2357

0.004 0.003 0.002 0.001 0.000

0.00187

BASE 1 000
Figure 5.3

. 0419 0 0 . 01 0 .0 ~ ~ e
. .

ALCOHOL NON PREFERRING


. .

ALCOHOL PREFERRING * P < .01

0.0019 *

0.00058

0.00042

500 1500 NUMBER OFLEARNING TRIALS

2000

-preferring -non oflearning inalcohol and rin~ networks. Comparison prefer

smaller than that causedby opiate antagonist treatment alone (z = - 2.80; P < .01). Combined treatment with opiate antagonist and serotonin reuptake blockers was significantly better than either treatment alone (opiate antagonist alone vs. opiate antagonist with serotonin reuptakeblocker: z = - 2.24, P < .05; serotonin reuptakeblocker vs. combination: z = - 2.80, P < .01).

DISCUSSION
The results of this study show that a computational model can simulate . Differencesin desireto drink alcohol were symptoms of alcohol dependence under both normal and abnormal conditions. Normal levels of produced desire to drink alcohol were consistently associatedwith inputs to the networks . Mild desire to drink was associatedwith weak stimuli, and strong desire was only associatedwith strong stimuli. In contrast , when the network changed to simulate the increasedlevels of dopaminergic function in the NAcc found in alcohol dependence , even weak stimuli elicited strong desire to drink. This finding is similar to the experienceof persons who are dependenton alcohol. Minimal environmental stimuli, such as seeing others drinking, and even internal stimuli, such as thinking about alcohol, can elicit strong desiresto drink. This condition replicatesthe situation hypothesized by Modellet al. (1990), when an alcohol-dependent individual takes his or her first drink.

135

Ownby : A Computational Model of Alcohol Dependence

Modell and colleagues argue that the acute effects of alcohol intake increasedopaminergic activity in the NAcc. This increasedactivity in turn results in higher levels of desire to drink via activity in the frontothalamic , the excitatory feedbackto the system from the fronloop . In this simulation to thalamiccircuit is represented activation of dopaminergicbias by increased units. The ability of the network to portray both normal and abnormal levels of desire to drink is a necessaryfirst step toward validating the model for other uses . Its ability to perform this simulation is not surprising in light of the capacity of computational models to simulate other psychiatric syndromes -Schreiber , such as schizophrenia(Cohen & Servan , 1992; Hoffman, 1987;) or anxiety disorders(Ownby & Carmin, 1995). The model was next modified to simulate a more dynamic process involved in the development of alcohol dependence : the acquisition of increased network activation in responseto environmental and internal cues . By increasingthe inherent dopaminergic activity of the loop, including the NAcc and VT A , it was possible to simulate genetically baseddifferencesin strains of rats. Alcohol -preferring and -nonpreferring rats have been shown to have different dopaminergic responsesin the NAcc and VT A to alcohol administration (Welss et al., 1993), with alcohol-preferring rats having larger increasesin dopaminergic activity after alcohol intake. The ability of the model to simulate this effect suggeststhat it could test any strategy that , if a might be developed for reducing this inherent activity . For example dopamine blocking agent selective for receptors in this circuit were to be , its effects on the development of alcohol dependencecould be developed assessed with this model. The network also was able to simulatethe effectsof both psychosocialand . As modeled , both categories pharmacologicaltreatments of alcohol dependence of treatment reduced abnormal levels of desire to drink, although psychosocial treatment was less effective than were serotonergic or opioi . The model predicts that combinationsof treatments , especially dergic treatments combinations of pharmacological and psychosocial treatment , are likely to be more effective than either treatment alone. Researchhas provided limited support for this prediction (Mason, 1994). The model also shows that augmentation of opioidergic treatment with serotonergic treatment should also be significantly better than either treatment alone. Preliminary data from an ongoing trial of the opioid antagonist nalmefene augmentedwith the serotonin reuptakeblocker sertralinesuggestthat this is also true (B. J. Mason, personalcommunication , May 8, 1997). The demonstration that psychosocial and pharmacological treatments result in similar effects provides an integrative rationale for both types of ' treatments effectson the brain. This is reminiscentof brain imaging findings in the treatment of OCD . Those studieshave shown that successful symptomatic relief of OCD symptoms by either pharmacologicaland behavioral treatment reducedmetabolic activity in a part of the basalganglia (Baxter et al., 1992). Thesebasalganglia structuresare believed by Modellet al. (1990)

BehavioralStates

to be part of the samecircuit involved in obsessionsand in the obsession ' like quality of alcoholics intense desiresto consumealcohol. Behavioral and pharmacologicaltreatmentsmay have impacts on different parts of the network , but have similar results on the samesymptom-related structures . The model used a partially recurrent architecturenormally used for problem . In solving in situations in which changesacrosstime must be learned this simulation , however, the network architectureis used to simulate neural loops that regulate the flow of information &om input to output units. The use fulnessof this model in simulating observed clinical phenomenais fulness will lie in its ability to test strategies intriguing , but its eventual use for changing the activity of this neural system and to generate testable hypothesesabout modes of treatment. The prediction that augmentation of opioidergic treatment with serotonergictreatment was made independentof fulness of sertraline augmentation of nalmefene knowledge about the use treatment for alcohol dependence and was a minor test of the model. Since the idea that addressing intense desire to drink via two neurotransmitter systemswould be better than via one is intuitive , however, this finding cannot be considereda strong test of the model. In the future, it is hoped that other tests of the model can be devised which are stronger. For example , it might be possible to compare the network ' s predictions on the relative efficaciesof several psychosocial treatments with actual treatment data. Would treatment focusedon reducing the salienceof alcohol-related cues be superior to that focused on coping skills -related treatment or relapse prevention? It could be hypothesized that cue would reducesensory and motivational inputs to the network, while coping skills might increaseinhibition of the circuit through parts of the &ontal cortex. Although the net effect of these interventions should be to reduce abnormally intense desire to drink, it is not clear which strategy might . Such a simulation is complicated by difficulties in produce larger effects the quantifying changes in different neurotransmitter systems caused by interventions. At present , only qualitative changescan be reliably shown, but it is hoped that further researchabout receptor binding affinities will allow at least rough estimatesof the relative effectson neurotransmissionof various opioid antagonists compared , for example , to serotonin reuptake blockers . Further development of neural modeling software and computa tional algorithms may solve this and other technicalproblems , and additional neuroscience researchon the phenomenonof alcohol dependencemay provide additional data with which to refine the simulation. Continuing work on the model thus appearswarranted.

ACKNOWLEDGMENT
This chapter received the 1996 ResidentResearch Award sponsoredby the American PsychiatricAssociation and Lilly Pharmaceuticals .

137

: A Computational Modelof AlcoholDependence Ownby

REFEREN CFS
Annals of alcoholism . Psychiatric in thepharmacotherapy , 25 , Anton , R. F. (1995 ). New directions - 362 . 353 .), :R . B. Millman(Eds . In J. H. Lowinson of craving Bauer , P. Ruiz ,& , LO . (1992 ). Psydtobiology . : Williams& : Wilkins : A comprehensive textbook abuse Substance (pp. 51- 55). Baltimore . S., Szuba M. P., Guze Baxter , J. C., Alazraki , , B. H., Mazziotta ,K , J. M., Bergman , L R., Schwartz rate metabolic E. . H. P. & : M. 1992 Caudate A., Selin Munford . , , , C. E., Femig , ( ) glucose , Phelps . Archives disorder -compulsive with both drug and behaviortherapyfor obsessive of changes - 689. General , 49 , 681 Psychiatry mechanisms of ethanolintoxication Chamess , and physical , tolerance , M. E. (1992 ). Molecular K. Me Do Eds . and treatment In . H. & : N. . Mendelson ( ), of akoholism Diagnosis dependence J - 199 -HiD . ). NewYork: McGraw (pp. 155 -Schreiber : A connectionist : Servan Cohen , cortex . and dopamine , D. (1992 ). Context , J. D. & . . Psychological Review andbiologyin schizophrenia to behavior , 99 , 45- 77 approach is neurotransmission . A. (1993 : Wise Devine ,R , P., & ). Mesolimbic , D. P., Leone dopamine . European , of p- opioidreceptor increased Journal of Pharmacology by administration antagonists 243 , 55- 64. : A neurobiological E., & i , E. (1992 : Carbon Di Chiara , G., Acquas ). Drug motivationandabuse - 219 . 654 207 New York Sciences . Annals the , , Academy of of perspective increase Di Chiara : Imperato , A. (1988a , G., & ). Drugsabused by humans synaptic preferentially . Proceedings in the mesolimbic concentrations of the systemof freely moving rats dopamine 5278 . 85 5274 National Sciences , , Academy of on of mu and kappa Di Chiara : Imperato , A. (1988b ). Oppositeeffects , G., & opiateagonists rats . of in the dorsal caudate ens and in the nucleus accumb release Journal freely moving dopamine - 1080 . andE , 244 , 1067 of Pharmacology Therapeutics :rptrimental involvedin ethanol channels in neuronal calcium Dolin . J., & : Uttle, H. J. (1989 ). Are changes ,S - 989 . andErptrimental ?Journal tolerance , 250 , 985 of Pharmacology Therapeutics : Little Dolin ton , J. (1987 , M., Pagonis , C., & ). Increased , SUttle , H., Hudspith dihydropyridine . Neuro in rat brain may underlieethanolphysicaldependence sensitivecalciumchannels . 26 275 279 , , pharmacology : Pomerleau . E. Corry . F., Meyer , O. F. (1987 , P., & ) ,R ,R , D. E., Kaplan , Z. 5., Morse Dolinsky in male Neuroendocrine , and subjective , psychophysiological reactivityto an alcoholplacebo . Research - 300 : Clinical andE . Alcoholism , 11 , 296 alcoholic :rptrimental patients on serotonin andclinicalstudies ., & : Unnolia , alcohol , M. (1992 ). Basic ,K GeorgeD. T., Wozniak interventions .), Novel : EM . Sellers . In C. A Naranjo& andalcoholics (Eds for pharmacologic -Verlag . : Springer akoholism ). NewYork (pp. 92- 104 . Reading to thetheory . G. (1991 : Palmer ,R , A., & ). Introduction Hertz , J., Krogh of neural computation -Wesley . , MA: Addison and the schizophrenia of neuralinformation simulations . E . (1987 Hoffman ,R ) . Computer processing - 188 -mania . . Archives , 178 , 44 Psychiatry of General dichotomy in the nucleus of dopamine release stimulation : Di Chiara , G. (1986 , A., & ). Preferential Imperato Therapeutics and E . Journal accumb ens of freely moving rats by ethanol of Pharmacology :rptrimental . - 228 , 239 , 219 . Annals of drugreinforcement Koob , G. F. (1992a ). Neuralmemanisms of theNewYorkAcademy - 191 . , 654 , 171 of Sciences

138

BehavioralStates

Koob : Anatomy , G. F. (1992b andfunction of reward ). Drugsof abuse , pharmacology . pathways - 184 in Pharmacological Trends Sciences . , 13 , 177 Koob : Bloom , G. F., & , F. E. (1988 mechanisms of drug dependence . ). Cellularand molecular - 723 Science . , 242 , 715 Koob : Welss , G. F., Rassnick , 5., Heinrichs , S., & , F. (1994 , the rewardsystemand ). Alcohol - 114 . EXS . , 71 , 103 dependence : Benkelfat and alcoholintake , D., Pihl R. 0 ., & , C. (1994 ). Serotonin , abuse , and LeMarquand : Clinicalevidence . Biological . - 337 , 36 , 326 dependence Psychiatry Lhuintre , J. P., Daoust , M., Moore , N. E., Chretien , D., Saligaut , C., Tran : , G., Bosimare , F., & Hillemand bis acetylhomo taurine , B. (1985 ). Ability of calcium , a GABA agonist , to prevent - 1016 in weaned alcoholics . Lancet . , I , 1014 relapse - 17 of craving . Alcohol Health andResearch , A M. (1986 World ). Themystery , 10 , 12 , 69. Ludwig Marlatt : Gordon . (Eds .) (1985 : Maintenance , G. A., & , J. R in the ). Relapse prevention strategies treatment behavior . NewYork: Guilford .. of addictive Mason . (1994 of psychosocial andpharmacologic treatments , B. J andcombined ). Effects individually on alcoholdependence . Presented at the annual meetingof the Societyfor Biomedical in Alcoholism Research 1994 . , Brisbane , Australia , June McBride . M., Lumeng : Li, T.-K. (1992 , W. J., Murphy and alcoholconsumption ,J ). Serotonin , L., & . In C. A. Naranjo & : E. M. Sellers .), Novel interventions (Eds pharmacologic for alcoholism -Verlag . (pp. 59- 67). NewYork: Springer Modell : Mountz andcompulsive , J. G., Glaser , F. B., Cyr, L., & , J. M. (1992a characteristics ). Obsessive of cravingfor alcoholin alcoholabuse anddependence . Alcoholism : Clinical andExperimental - 274 Research . , 16 , 272 Modell : Lee of haloperidol , J. G., Glaser on measures , F. B., Mountz , J. M., & , J. Y. (1993 ). Effect of craving and impairedcontrol in alcoholicsubjects . Alcoholism : Clinicaland Erperimental - 240 Research . , 17 , 234 Modell : Cyr L. (1992b , J. G., Glaser , F. B., Mountz , J. M., Schmaltz andcompulsive , 5., & ). Obsessive characteristics of alcoholabuse and dependence : Quantification by a newly developed - 271 . Alcoholism : Clinical andExperimental Research . , 16 , 266 questionnaire Modell . M., & : Beresford , J. G., Mountz ,J , T. P. (1990 ). Basal / limbicstriatalandthalamo ganglia corticalinvolvement in cravingandlossof controlin alcoholism . Journal and of Neuropsychiatry - 144 Clinical Neuroscience . , 2, 123 Modell : Greden . F. (1989 , J. G., Mountz , J. M., Curtis , G. C., & ,J ). Neurophysiologic dysfunction in basalganglia circuitsas a pathogenetic mechanism of / limbic striataland thalamocortical obsessive -compulsive disorder . Journal ,1 , 27- 36. of Neuropsychiatry Monti, P. M., Abrams . M., & : Cooney , D. B., Kadden ,R , N. L. (1989 ). Treatingalcoholdependence : A copingskillstrainingguide . NewYork : Guilford . . M., McBride of monoamines in forebrain ,J , W. J., Lumeng ). Contents Murphy , L., Li, T.-K. (1987 of alcohol P and non lines of rats . ( ) , Biochemistry ( NP ) regions preferring ring Pharmacology prefer - 392 . , andBehavior , 26 , 389 Nestler mechanisms of drug addiction . Journal , E. J. (1992 ). Molecular , 12 , 2439 of Neuroscience 2450 . ' O Malley . 5., Meyer . E., & : Rounsaville , S. 5., Jaffe , A. J., Change , G., Schottenfeld ,R ,R , B. and copingskills therapyfor alcoholdependence . Archives (1992 ). Naltrexone of General Psy - 887 . , 49 , 881 chiatry

139

Ownby : A Computational Model of Alcohol Dependence

. L., & : Carmin ,R networkmodelof panicdisorder , C. N. (1995 ). A neural Ownby (abstract ). Biological . , 37 , 616 Psychiatry Rohsenow ., Childress , D. J., Niaura , R. S : Monti, P. M. (1990 , A. R., Abrams , P. M., & - 1991 ). Cuereactivityin addictive : Theoretical behaviors andtreatment . International Journal implications - 993 . , 25 , 957 of theAddictions Rumelhart : Williams , O. E., Hinton , G. E., & internalrepresentations , R. J. (1986 ). Learning by error propagation . In O. Rumelhart Uand : the POPResearch , J. L. McCle ,& .) Parallel Group(Eds - 362 distributed : Explorations in themicrostructure , Vol. 1 (pp. 318 processing of cognition ). Cambridge : MIT Press . , MA
Sellers : Sobell , E. M ., Higgins, G. A ., & , M . B. (1992). 5-HT and alcohol abuse . Trends in Phanna co , 13 , 69- 75. logicalSciences : Castell an, N. J. (1987). Nonparametric , S., & statistics , 2nd ed.. New York: McGraw-Hill . Siegel Stinus : LeMoal , M . (1992). Interaction between endogenousopioids and dop, L., Cador, M ., & amine within the nucleusaccurnbens . Annals oftht New York Academy , 654 , 254- 273. of Sciences Terenius , L. (1994). Rewardand its control by dynorphin peptides . EXS , 71 , 9- 17. Trevisan , L., Fitzgerald , L. W., Brose , N., Gasic , G. P., Heinemann : Nestler, , S. F., Duman , R. S., & R . J. (1994). Chronic ethanol upregulatesNMDARI receptor subunit immunoreactivity in rat . Journalof Neurochemistry , 62, 1635- 1638. hippocampus Turkkan, J. S., McCaul, M . E., & : SHtzer , M . L. (1989). Psychophysiologicaleffects of alcohol related stimuli. II. Enhancement with alcohol availability. Alcoholism : Clinical and Experimental Research , 13 , 392- 398. : Johnson , K., & and installation , E. F. (1996). Linux , configuration Volkerding , P., Richard , 2nd ed.. New York: MIS Press . ' : O Brien, C. P. (1992). Naltrexone in the treatment , M ., & Volpicelli, J. R., Altennan, A . I., Harashida of alcohol dependence . Archivesof General , 49, 876- 880. Psychiatry Welss : Koob, G. F. (1993). Oral alcohol self-administration , F., Lorang , M . T., Bloom, F. E., & stimulates dopamine releasein the rat nucleus accumb ens: Genetic and motivational determi-

. Journal nants andExperimental - 258 . , 267 , 250 of Phllrm Rcology Therapeutics , M. A., Dolin , S. J., Patch , T. L., Siarey , R J., et al. (1991 Whittington ). Chronic dihydropyridine treatment canreverse the behavioural of and to chronic ethanol consequences prevent adaptations treatment . British - 1676 . , 103 , 1669 Journal of Phannacology -Hill. Winer in experimental , B. J. (1971 .. New York: McGraw ). Statistical , 2nded principles design Zell, A., Mamier : Vogt, M. (1996 , G., & user 's manual ). StuttgartNeuralNetwork Simulator , version 4.1. Stuttgart : Universityof StuttgartInstitutefor Parallel andDistributed High Performance . [ TheWorld WideWeb pageon the simulator is at http://www.informatik .unistuttgart Systems .dejipvr .html. The softwareis available for downloadat: ftp:jj / bvjprojektejsnnsjsnns .unistuttgart .dejpub .] / SNNS ftpinformatik

140

BehavioralStates

A Computational Perspediveon learned andDepression Helplessness


Sam Leven

One of the most persistent syndromes in human behavior is depression . It to , 1997) and partly to psychotherapy yields partly drug treatment (Richelson : Greenberger , 1997; Padesky& , 1995). More intractable is atypical (Beck , which is characterizedby a mixture of problematic behaviors depression associatedwith the underlying dysphoric state. Tyrer and Steinberg (1993) have suggestedfive mutually exclusive models of behavior dysfunction: disease , psychodynamic , behavioral , cognitive, and psychosocial (table 6.1). Each approach has its own internally consistent theoretical basis , treatment and evaluations . However each alone an , , options provides inadequate . explanation of depression In this chapter es that produce , we present an integrated view of the process a based on neural network architecture . Neural networks depression to be suitable for eminently modeling complex phenomenainvolving appear neurochemicaland neuroelectrical phenomena . There , such as depression is no other environment for asking such complex questions as , What role does inhibition of one sensory process play in some apparently unrelated ? What effectsdo changesin expectancies behavior and preparatory set have in speeding this behavior or magnifying its importance ? And , what is the " " , Adler , Dreyfuss , et al., 1988) on impact of the chemical signature (Black the representationalprocess ? Investigating such questions requires a branch of neural networks that comprehends higher-order process es (Grossberg , es must be reductionist, 1980). While neural network models of such process , to pose testable theories in a they remain a useful way to query the system , and to compareexplanations of what may not seemfully rigorous manner data , 1988). (Leven comparable ' " In the following section , we describe a model of leamed helplessness , 1975), a dysfunction frequently consideredan animal model for (Seligman human depression and anxiety. Despite ongoing controversy about the causes , course , and treatment of learnedhelplessness , 1992; Peterson , (Leven Maler, & : Seligman seemsto be forming that allows for , 1993), a consensus the formalizing function and modeling clarity neural networks can provide es some of the controversial (Levine, 1991, chapter 7). The next section address ' s research issues . In addition, the important assumptionsof Samson

foundational consensus , Mirin, Hauser (Samson providethe emerging , et al., . 1992 )

TYPFS OF HELPLFSSNESS
Theories of the causeand treatment of depressionare numerous and sometimes . Psychoanalytictherapies(Arieti & : Bemporad , 1982) focus inadequate on previous infliction of injuries to the self; cognitive therapies (Alford & : Beck , 1997; Beck , A ., 1976, 1997; Beck , J., 1995; Beck , Rush , Shaw , et al., 1987) find failures in explanatory styles; somatic therapies (Lowen, 1971) treat sensorimotor conflicts with the current environment. Among psycho, the conflict is not over therapies as much as theories pharmacologists : Roth, 1991): which brain structuresand functions account (Cooper, Bloom, & for the development of depression ? The animal models upon which human treatmentsusually rely have produced many theories . Leven (1992) has suggested some potential limitations in these models . We briefly summarize recent work by leading scholarswho bring together the experimental traditions of Welss , Seligman , Petty, and collaborators , and then review ours. 6.1 work from two separate experimental traditions Figure represents which converge toward a single solution. Minor , Dess , and Overmier (1991) have describedthree sets of experimentswhich induc~ three similar but differentiable . They describetheseas motoric, subtypesof learned helplessness and . 1987 advanced a related model. His three ) vegetative, cognitive Gray ( , but do not explain the broad systems systemshave biological components of neuromodulatorswhich underlie psychopathology (van der Kolk, 1987). In the human post-traumatic stressdisorder literature, a closely matching set of categories is introduced (for reviews, see Herman , 1992): hyperarousal , constriction, and intrusion. Thesemodels derive from different sources . Laboratory work in animal models has emulatedhuman behavior through shock , variable schedulesdesigned to discourage , and isolation. Lesions and other intrusive methods have allowed development of schematic models. Drug , 1997). therapies have suggested underlying chemical systems ( Richelson And human subjects (both depressive and anxiety disorder and traumatic stress sufferers ) have confirmed the outlines of the experimental observations . What remains is to create a synthesisof these efforts, design testable

142

BehavioralStates

Figure 6.1 Theoretical models of depressionand post-traumatic stressdisorder.

network architecturesto emulate them, and then redesign them as new data arise . The physiological circuitries of the three subtypes are described by all in varying detail, but with some agreement . Their chemical theseresearchers substratesare described in the experimental traditions we have surveyed above and elsewhere(Leven , 1993). The evidence suggests three ways of . Motoric (using , acting, and perhaps stopping being helpless becoming ' " " Minor , Dess : Overmier s terms , an inability to ,& ) depressionis a stuckness avoid repeatedand ineffective behaviors , or simply giving up becausethose . Its main chemical message are the only choices available , norepinephrine , intermittent frustration of simple , and main behavioral cause ( NE ) shortage " " behavior, have been resolved in laboratory experiment by radical increases in NE supply and high physical arousalleading to consistent success , ( Welss 1991).1 In human subjects , Herman (1992) finds chronic stress and an elevated " " baselineof arousalwhich lead to either form of stuckness observed in the animal literature: perseveration without recognition of novelty , or freezing. " " , lack of good feeling. Being separated Vegetative depression is aloneness from loved ones under arbitrary conditions and being deprived of " " endogenousopiates have produced this senseof personalloss. Eventually, , consistentbelonging with loved voluntary isolation tends to result. Induced ones and induction of supply of endogenousopiate resourcestend to restore . In humans effective behavior in laboratory animals , Herman (1992) describes

: Learned andDepression Leven Helplessness

INFERENCE DRAWN FROM LEARNED , ANTICIPATED EVENT

COGNITIVE : FAILURE AND ATTRIBUTION ANTICIPATION FAILURE

" " -Beck . Figure6.2 Seligman cognitive depression

alterations in pain tolerance , fear of painful social encounters , and withdrawal " self" with tend to medicate induced patients externally opiates and alcohol. Cognitive depression , described by Beck (1976, 1997) and Peterson et al. (1993), involves misuse of available information about the possibility ('f effective behavior (figure 6.2). The confusion can lead to fear and immobility . Dias, Robbins , and Roberts (1996) have confirmed the model presentedby Pribram and McGuinness (1975), that the samesystem which provides the basis for semantic categorization also facilitates attentional focus in frontal regions. , and Hendrickse (1993) have eliminated a similar deficit in Petty, Kramer laboratory animals by inducing greatery -aminobutyric (GABA ) availability

144

BehavioralStates

's in cortical regions; Beck cognitive therapy enables efficient analysis and accurate explanations of events that have led to depression (Beck , 1976, 1997; Beck , Rush , Shaw , et al., 1987). Like Beck , Herman (1992) finds that false or inadequateexplanations tend to induce reasoning from distorted " " grounds, and the need for a superior explanation to the one currently employed. While these types may be separatelycharacterizedand detected , they are . Instead , as in most complex behaviors , multiple unlikely to be totally unrelated " " feedbackloops (Miller , Galanter : Pribram ,& , 1960) lead to the very the clients practitioners encounter atypical, mixed problems which characterize leven , 1987) today ( These same systems emerge in autistic disorder and pervasive developmental -Cohen -Cohen suggeststhat a disorder (POD) (Baron , 1995). Baron senseof self required to develop intentionality and the ability to recognize ' others intentional statescan be interrupted at three preliminary stages(eye direction detection, intentionality detection, and development of a " shared " attention mechanism ). Failure at theseearlier stagesprevents induction of a " " ' of mind mechanism , which facilitates the recognition of one s own theory ' and others meaningfulbehaviors . Baron Cohen specifiesneurological centers which permit development of these facilities. Eye direction detection takes place in the superior temporal sulcus , intentionality detection in the temporal lobes , and shared attention in the hippocampalcomplex. The ultimate development of one' s awareness own " theory of mind," Baron-Cohen asserts , takes place orbitofrontally . These general functional mappings , drawn on researchon dysfunction, are also reflectedin theoretical analysesby Pribram (1991) and Maclean (1990). , Psycholinguistic analyses continue to support these views (Fauconnier 1994). Specificemotional mappings which are similar include work by Gray (1987), Gilbert (1984), and Leven (1992).

: MasterJ, V8. Control Samson


Our target problem is a finding of Samsonet al. (1992): There is a link betweencognitive process es associated with coping responses and noradrenergic neuronal activity . . . . [ We ] suggest the possibility that noradrenergic neuronal activity is related to the cognitive interpretation constructed by the individual to explain environmental events higher is associated with of , and norepinephrineoutput perceptions powerlessness lower norepinephrine output is associated with perceptions of mastery. (p. 808) ' Samson s work identified two statistically distinguishablegroups of helpless (depressed , distinguishing them by their neurochemicalstructure and ) people the way they describedtheir senseof helplessness (explanatory style). ' The distinction in subjects discussionswas based on a pair of standard instruments for measuring whether people feel that (1) they control the

145

Leven : LearnedHelplessness and Depression

essentialdecisionsand actions they take; (2) they act at random- there is no control they can detect over their own behavior; or, (3) other forces control their behavior. The three groups are plainly separableand their views are essentialto the therapeutic approach . The cognitive therapy of depression Beck 1976 1997 Beck Rush Shaw , , ( , , , , et al., 1987; Petersonet al., 1993) suggests that depressivesfall into two groups: those with purely organic difficulties (such as brain injury ) and those who have faulty explanatory style. The group that lacksphysical impairment is held to mistakeunfortunate outcomes with failures they have caused . Treatment consistsof providing accurate information to eliminate the inferential failures (while curing problems that have resultedfrom the inappropriate thinking ). Samsonet al. (1992) produced a findi11g that presentsa problem for traditional models of depressionand its animal analog . The , learnedhelplessness group, following the leading work of Schildkraut (1965) on the catechol amine theory of depression , has established the unique role of NE in a of behavior , et al., 1989). The most recent (Shatzberg subtype depressive of Samson and established that human subjects testing findings colleagues low on a marker for NE level (the metabolite methoxyhydroxyphenylglycol ) also experiencedan increasedsensethat outside influencescontrolled outcomes in their lives. NE provides the " focusing" processin brain function (Willner , 1985)2 and 's it is generally consistent with Grossberg gain control function. Further, it inhibits serotonin (5 HT ) releasefrom the amygdala , interrupting data on " " values (Pribram , 1991)- in some cases , the releaseof high amounts of NE can block releaseof effective quantities of 5-HT , resulting in deletion from the system of critical information about rates of success , identification of relevant positive information (through interruption of its feedbackto cortical " " areas ), and preventing the blocking ( forgetting ) of traumatic information (Deakin, 1994). NE enervation also facilitates effectivenessof acetylcholine ; this allows motivational featuresto be delivered (A Ch) releaseand reception throughout the brain. The main sourcesof NE are the basal ganglia and hypothalamus ; these " " regions feed the locus ceruleus(LC), the central source for the LC system , 1985). Deakin (1994) reports that the median raphe nucleus (Nieuwenhuys , a serotonergic body placed in the center of the information flow , plays the unique informational role of identifying learned stress stimuli, sources of chronic fear, and environments in which control is absent . A signal to the allows that to stressful recollections (preventing amygdala organ interrupt " " ' phenomenalike post-traumatic stressand freezing, as argued in Hermans, 1992, review). The dorsal raphe nucleus is the source of information on future threats and anticipations of anxiety-producing circumstances in cortical 'learned " Deakin . In addition a 1994 recent , , ( ) regions helplessness positron emission tomography study suggested reciprocal diencephalic and limbic activation with neuropsychological tests of solvable and unsolvable , Gur, Alavi , et al., 1996). anagrams(Schneider

146

BehavioralStates

We interpret Samsonto suggest , as Welss (1991) does , that in the presence of sufficient NE, motor and self-control- related helplessness would not , even if experimentalsubjectswere placed in variable-schedule , painemerge . Rat subjects , Welss and Simson (1989) have reported, have producing tasks resilient to such . These subjectswould be consistent proved discouragement " -produced depression with traditional cognitive" explanationsof inference .J Yet the NE reducedhelplessness Samsonobservedcould only happen in the -style situation (uncontrollable stress Welss accompaniedby exhaustion of NE production at the locus ceruleus ). This animal experiment result would 's tend to confirm Samson finding that those with chronically exhaustedNE ' resources would act like Weiss s NE-reduced rats. In other words, they would show someevidencethat their own actions would have no effect and that others would control their worlds.
Norepinephrine Dopamine and the Loss of Internal Control Without Explanation :

Samson et al. ( 1992 ) established a clear marker of a specifically detectable depressive symptom . It was the stable coincidence of a specific chemical state (low NE utilization ) with a specific dispositional state (loss of sense of personal control ). This detected symptom was the way subjects explained their beliefs about their own condition . Hence this was a higher -order marker . Other symptoms , when related through the same discriminant analysis , showed no similar statistical significance . Among these unrelated symptoms were other explanations . These other nonsignificant explanations have been related , in the experimental literature , to other neurotransmitters . This result " " alters the single -bullet approach to theories of depressive mentation involving chemical and dispositional state. Samson and associates showed that a separate symptom was tied to a subtype of behavior . We expand on their foundation and suggest the symptom could be causally linked to the behavior by existing theory and observation . The contrast in the two groups would have been clarified had Samson et al. ( 1992 ) established whether homovanillic acid (HV A ) levels were low in their other group . HV A is a metabolite of dopamine (DA ); a diminished level might suggest that the raphe-potentiating signal to cortical GABA was the source of the alternative behavior .4 This would have further emphasized the qualitative distinction between two architectures , motivated by different chemical message systems, determining different types of depressive behaviors . Had the two groups been differentiated on that basis, two types could have been understood on the basis of an existing typology of both learned ' helplessness and depression (Leven , 1992 , 1993 ). The chemistry of Weiss s ' and the Samson group s helplessness, NE and 5 -HT , could have been distinguished " " &om the cognitive type suggested by work by Deakin ( 1994 ) that we and others have discussed, produced by a different physiological route and different neurochemistry .

147

Leven : LearnedHelplessness and Depression

ENDOGENOUS OPIATES
DRIVE > . : / Af IDI K F ' ' Jf rIID W J . . Ar or It A v V . < M An CNZSSIS DIIVD D DI8NDS A QUA ' ~ ; Am UrDmIU Wrl SrsJD f ALARM

PLANNING ALERT
: Nrx A M I f .

-> FAlTB < DESPAIR


lmrl , ' : ' . S 1 , N ON I T1 I ' O Jr D0 I Ar Wrr Ifm Im T Atm Sl Nr S I D U N A " Iocas I Im 1 CO S A GlSTAm Nl 1 I SE A I O

-BECK SELIGMAN

Figure 6.3 A systemsview of dysphoria and post-traumatic stressdisorder. The model as initially conceived .

' " " A Model of Samson s Mastery Type How can the same stressful, discouraging event produce two qualitatively different types of helpless behavior ? First , we describe our hypothesis in some detail . Then , we report an experimental result which clarifies further

. the distinction we have drawn and raiseshigher-order questions In order to understand our hypothesis , it is helpful to examine in more detail the network diagramsof figures 6.3 and 6.4. Note that, in both figures , the dorsal raphe recognizesthe incoming stressand, eventually, producesa DA output to cortex. In figure 6.3, this output (becauseit does not suggest impending failure) does not reach threshold quantities at cortex and is

148

BehavioralStates

AL

It

Ar

ma lM Z 01 ) DO If Nr l D I I MS ml . N

1J rm I mlmJlm RI8AICIIK rAJ

SELIGMAN (1975 )

WEISS

FREEZING

PAIN

' s model). Pigure6.4 Stressproducesfreezing ( Weiss

ignored. Meanwhile, the median raphe and LC observe the stressfulsituation " and fire an identifying " signature signal electrically (from the dorsal raphe ) and a stressfulNE signal from the both median raphe and LC to the hippo. The stresssignal inducesheightened plasticity at the hippocampus campus for the raphe signal and excites 1 .-adrenergicreceptors , Welss (1991) asserts , which mediate NE activity here. However, the plasticity is insufficient to induce storage of the signal. The signal then falls below normal and the 1 .the shortfall , sensing , upregulate (increasetheir population density) receptors

149

Leven : LearnedHelplessness and Depression

at the reception site as a normal part of homeostasis. Eventually , NE stocks are replenished , although at a lower level than normal . Ordinarily , the receptor population would diminish . Instead, in the Welss training , helplessnessinducing stress is reintroduced before the drop in receptor population . The rush of NE again reaches the hippo campus with the raphe signal , then again falls to below normal levels . Upregulation of !Xl-receptors takes place, further increasing the receptor population . Finally , after return to somewhat subnormal NE levels , a third Welss training occurs. The rush of NE sensitizes an even larger number of !Xl- receptors to the signal , finally bringing the signal above the threshold required for memory impression . " " Now that the signal is imprinted (Bateson, 1972 ); an automatic response is engendered to the stimulus . The state of helplessness is recalled , and the shortage of NE is experienced . The receptor population , now returned to a " " near- normal state, experiences none of the driving influence of NE and yields an indication to cortical levels that failure has occurred .

Affect-Driven

GADA

Let us turn now to figure 6.1, which presentsthe contrary case . The dorsal identifies a future threat and releases sufficient DA to impress raphe potential the cortex that something significantly stressful is about to occur. Information presented to the cortex confirms that failure has occurred (Levine, Leven : Prueitt, 1992). The chemicaland informational signal suggeststhe ,& predictive capacity of this event for failure. As a series of conditioning models demonstrates , repeatedpresentationof co-occurring signalswill produce " " , 1980). Our model represents cognitive learning (for review, seeGrossberg this mechanism by sending a dual signal from the dorsal raphe to the neocortex. The informational (electrical ) pattern denotes the actual anticipated circumstance and the DA records the urgency. Petty et al. (1993) and Suomi (1991) have suggestedthat the DA signal is mediatedby GABA in neocorticalregions (figure 6.S). GABA inhibits frontal " " noise (Cooper et al., 1991) and enables focus on relevant information which would distinguish most situations from the recalled or " trained" case . This work suggests(Leven , 1992, 1993) that reduced GABA , produced as a result of the DA signal , could lessenthe ability of the cortex to distinguish -producing and non- helplessness -producing behaviors between helplessness ' " " (reduce Grossbergs gain control ). The resulting confusion could lead to ' overgeneralization of failing cures and more directly produce Samsons " " cognitive helpless group. Perseveration , Automaticity , and Norepinephrine - Mediated
Depression

' , Fennan , et al., Experimental results from Fieves group (Corwin, Peselow " 1990) support the notion that NE mediation of " decision es process among

150

BehavioralStates

CORTICOLIMBIC RECALLED SADNESS

.. , , , , , , , , , , , , : . . , S , -BT , : , :

' NJ ,, : NE , : " . '' . : : , " " " " ," " " ." " . . \ . .. . :.. MEDIAN RAAiE ' .. . . .. LEARNED AVERSION

. based on a synthesis of the ideasof Sumni Figure 6.5 Schematicof withdrawal depression

etal . (1993 (1991 ) and ). Petty

Leven : Learned and Depression Helplessness

differs from other chemicalsignaling effects . The experimenters depressives presented exemplars and possible matches to a group of depressivesand 's varied the reward schedulefor successful . Fieve categorization group tested them, then treated one group with propranolol (a blocker of NE). While all groups of depressivestend to be cautious about category identification and " " switching, the NE-deprived depressivesshowed decision rule deficits. In other words, they employed conservative , less risky strategies for categorizing items about which they were uncertain . The NE-depressedgroup madeno more errors in discriminating between " right " and " wrong " answers than the others. NE-deprived subjectswere less flexible, more automatic in deciding in an uncertain situation. The chemical impairment led them to , despite adverse consequences greater consistency of choice (Levine et al., 1992). These choice blases are unrelated to the better-known "heuristics" : Tversky, 1982), which involves systematic adjustments to (Kahneman & information. This pattern of systematic error, normally attributed to cognitive deficit , 1976, 1997), may be differentiated from other errors which require (Beck . The subjects are more averse to risk complex categorization and analysis because of traumatic experience(as in the post-traumatic studies summarized in Herman , 1992 , chapter 2). The stable failure is predicted by stressof the -Samsontype, not by attributional and Welss . This result expectancyeffects mimics frontal lesionswhich lead to perseverationin formerly rewarding behavior : Geder (Es : Benton, 1991; , 1996; Levin, Eisenberg linger, Grattan, & , & Levine, Parks : Prueitt, 1993). Both the perseverationin failed efforts and ,& the delays in responding to changeare present in the Fieve group results . Hence NE deficits which are often , induced by a depressive state , , can produce failures that have been categorized as cognitive. These results underline the complexity of the depressivestate and the profundity of the " " catecholaminebalance hypothesis. This state will be compared with its counterpart in the appendix. There, we illustrate the differing anatomiesand chemistries of these two conditions and describe simulations which qualitatively capturetheir unique qualities. A COGNIn V E SYSTEMSINTERPRET A nON OF DEPRFSSION This preliminary model follows the models laid down by Levine and Leven (1992) and Grossberg (1980). We have emulated available descriptions of neural systems , including neurochemicalsignaling (Black et aI., 1988). This model allows for chemical , physiological (lesion), and behavioral inductions and eliminations of distinct subtypes of helpless behavior in animals . As Samson(personal communication , September25, 1995) and Herman (1992) -and-fear systems are reentrant , these stress . That is, a cognitive emphasize form of induced depression('1 am logically bound to fail" ) and the resulting unsatisfactory outcome leads to learned motor and automatic system re-

152

BehavioralStates

, Welss depression ). Such a set of discouraging sponses(a stimulus response outcomescan produce low enough self-esteemto produce isolation and concomitant affective depression(Suomi , 1991). ' We adapt aspects of Grossbergs (1980) notion of intuitive mismatch , where largely insulated distributed systems require understanding of other their significanceoften comesat their interaction. Our such systems , because model placesimportance on cognition, conation, and affect. We have made significant progress by applying neural network principles to successive approximations of human behavior as convergent themes across neuropsychology and modeling emerge (Hestenes , 1992, 1993; , 1992; Leven Levine 1991). Pribram and Gill (1976) remind us of the task, saying that " there is a place in the scientific schemefor investigators and practitioners working at the interfacebetween disciplines [and] often, though not always, the most significant advancesin understandingand in practice arise at such " . 168 . interfaces ) Experimentalistslike Samsonand Schildkraut advance (p models and perform experimentsto establishconnectionsbetween behavior, . As theorists, we hope to build upon and , and therapeuticsuccess symptoms advancetheir efforts.

, HELPLFSSNESS - IN A MODELOF LEARNED : BUILDING APPENDIX STEPS


" " Step 1: The American Psychological Society Model in Principle Beginning with the models presented above, we specify the central biological units involved. We recognizethat three parallel systemsare required . This insight (first discussedin to emulatethe three subtypesof helplessness the much clearer after became Leven , 1992), appearanceof the article by Minor and associates (1991). A formal model that could capture the qualities es seemedto promise a way of designing described in the two approach . and biological experiments computational For presentationat the American PsychologicalSociety (Leven , 1993), our were charac simulation simply consistedof four interacting networks. These terized by their apparent similarity in processing output to the three biological networks and a trivial model of frontal function. Little effort was made to emulate critical upstream functions. The goal was to demonstrate " " the feasibility of producing three characteristic disabilities , using a set of . networks with different topologies and underlying theories This system , broadly described in figures 6.4 and 6.6 (the latter to be described below under Step 4), is very similar to the dialogic neural networks describedin Leven (1988) and Leven and Elsberry (1990). As in those sources , the networks were classicbackpropagation (of the Werbos, 1974, ) type representingsemanticfailure; the standardHopfield Tank (1986) model representing motoric failure; and standard adaptive resonance(of the Carpenter and Grossberg , 1987, type) representingaffective failure. The frontal

153

and Depression Leven : LearnedHelplessness

lobes emulation was a summation operator , allowing failures to be represented as outputs of the system . While it was possible to demonstrate the qualitative behavior of the described systems, it was impossible to demonstrate interactive effects. Such a model was adequate only as a demonstration in principle .

' Step 2: The World Congress on Neural Networks 95 Model The next step was to study the computational possibilities of a model which processedsemanticinformation in three qualitatively different ways and to relate those to a model of depression . We recognized limitations inherent in the original Leven (1988) choices : standard backpropagation " lacked the capacity to build extensive categories or " dynamic schemas (Rumelhart , Smolensky , McClelland, 8t Hinton , 1986) and the Hopfield-Tank model lacked learning ability and a reasonable means of modeling chemical signaling. We modified backpropagation(as discussedin Leven , 1995) by training simple systems to recognize three classesof logic: default (if -then), modal (when or how-then), and nonmonotonic (if pattern-then). These networks and categories(figures 6.7 and 6.8). proved capableof parsing classifications we trained basic backpropagationnetworks to recognize basic Similarly, set relations: into (included in), onto (equivalent to ), and subsumes (includes these ). Thesenetworks recognizedhierarchicalclasses 6.9 . (figure ) , we were able to design semanticnetworks capableof Combining these basic classification and logic-tree construction. We devised a simple mismatch " control structure which was designed to allow the system to " reset when it receivederror feedbackfrom other subsystems , motoric, or (affective frontal lobes ), as shown in figure 6.9. The combination of capabilities allowed us to claim that we had developed a continuously learning semantic system (figure 6.10). The other focus was to revise the motoric system , to allow it to employ the limited flexibility that human procedural systems allow. The vector Preisachmodel (Mayergoyz, 1989) had the capacity to accommodatemany , and modify learning in the inputs, store associatedpatterns within clusters " face of failure of its "base structures (minima). We suggesteda means of . modeling the NE systemusing this approach

Step 3: Washington Evolutionary

Systems

' Society 95

By now, we had reviewed the literature more carefully and had begun to describe individual structures . This was the beginning of developing the capacity to test experimental data against the model structure. As Levine (personalcommunication , Jannary23, 1993) has suggested , the hippocampus seems well analogized by adaptive resonance - like architectures . Dipole -filtering and merging functions that fields seem best suited to the sensory

BehavioralStates

Expectancycues mental spaces and creates frame: Each Modality (Habit, Affect, Semantic ) respondsto stimuli; Modalitiesare blendedand the blend is attachedto a cued frame. Consonant 's view of " " " " with Pribram deep and emergent structures.

FRAME

Figure 6.6

" Overview of the " work in progress model: seeappendix for details.

lSS

Leven : LearnedHelplessness and Depression

SEMANTIC INFORMATION PROCESSING

. DATA INPUT AND : STORED ; ; IN CATEGORICAL STORAGE CATEGORYBY COMPARISON WITH TRAlNm . TEMPLATE NETWORKS L T ERROR

COMPARISON OF ERRORS BDISTANCE FROM OPTIMUM


I MODAL LOGIC FIT I I OEFAULT LOGIC FIT I

o N ~ g ~ ~ l1I INO
1 1 1 1

t 1 1 1

1 1 1 I DATA STREAM

based on idealdatatypes . Figure6.7 Inference underlie it and to represent interactions between limbic and output areas. These dipole fields were introduced and seemed to replicate the reported functions (figure 6.11). Step 4 : Ongoing Work

Recent work by Self , Barnhart , Lehman , et al. (1996) points to increasing . 01 receptors in corticolimbic systems complexity in describing these systems induce different responses from 02 receptors in the presenceof stimulants . 01 agonists couple attention to, and excitement about, novel events with pleasure ; 02 agonists induce craving for excitation. Self is concerned with addictive responses to 02 stimulation; we may generalizethat 02 sen sitization producescraving for excitation, even in normals . The simpler role of 01 , producing focus and pleasure the basic , suggests deprivation model we produced earlier . We are attempting to , but combined with 02 responses . Rebound may be mediated by produce a model which has rebound effects in some normals and opiate production may be minimized in others (see 6.6 and 6.12 . This is also in a preliminary model being reflected ) figures

156

! BehavioralState

.... - . -

. by errorsin dataandstructure Figure6.8 Recategorizing

HIERARCHICAL SEMANTIC PROCESSING

CA1E8ORIE8 rAN BECREATED CATEGORY , BUILDING MERGED , ORStft Rm Dm BY f A8PA RlSO N WITH our PUr FRc . TRAINED ND- . K8

" I

INTO

"

FIT

. I r

ONTO

. FIT

INC LU D E '"FIT I " I

! ! ! ! ! ! ! ! ! ! ! ! "' - - . .'...-..'P ~ ~ ~ ~ .. - - .~ ~ ... ~ ~ .:: : t :: ~ SCHEMAS WEIGHT IETI


Figure 6.9 Hierardtical process es , derived from ideal types.

157

Leven : LearnedHelplessness and Depression

-FRONTAL DORSOLATERAL PRE CORTEX : LOGICALMEMORY OFTHEFUTURE

. II ~ = : : : -:~: : : : I

LOGICAL CLASS LAYER : MERGING OF SCHEMAS

. FROM DAT TO RUL : MAKI BUIL HIERA FRO INP TO CATE TO SCHE TO LOG CLAS RUL 1111 1111 1111 1111 1111
of categorization into semantic nets . Figure6.10 Integration

. Figure6.11 The modem model of affective processing

. " . ' . ' . ' . ' .

" .

"

: ~ !

~ ~ ~

fa i ~

'

tc

, !

, i

~ ~ ~~

o ~ ~ " . ~ : : / ) ~ ~ ~ ~ ' " ' ' . 4l1

./) "..
8 "li1 '

. In addition , testablemodelsconcerning by Pribram developed receptor issues are addressed . specificity currentlybeing

NOTFS
1. Among the many abnormalities induced by Welss and in related work are disturbancesin , appetite , weight, and arousability by changesin environment or novelty . sleep 2. NE is generally inhibitory (Cooper et aI., 1991, chapter6). As do all central neurotransmitters , NE perfonns a variety of aitical functions at different locations in the brain ; these are beyond the scopeof this discussion . 3. Samson(personalcommunication , September25, 1993) points out that multiple devicesmay result in an NE-mediateddepression ; heightenedNE turnover and functional deficits in the system are both indicated . 4. There might be three possible behaviors in the Samsongroup, not two (see Leven , 1992 ; Minor et aI., 1991 ). If the other group were at least partly homogeneous , HV A levels might . In that case , on might also study endogenousopiate levels, as suggestedby prove ambiguous Maler (in Petersonet aI., 1993) and draw finer distinctions among the companiongroup.

REFERENCES Alford, B. A., & : Beck . A. T. (1997 ). The . New York : Guilford integrative power of cognitive therapy Press . Arieti, S., & : Bemporad andmilddepression , J. (1982 . NewYork ). Severe : Basic Books . -Cohen Baron : An essay , S. (1995 onautism ). Mindblindness andtheory . Cambridge : , MA of mind MIT Press . Bateson toward anecology , G. (1972 . NewYork ). Steps : Ballantine Books . of mind andtheemotional Beck . A. T. (1976 ). Cognitive disorders . NewYork: Meridian . therapy Beck . A. T. (1997 ). Thepastandfutureof cognitivetherapy . Journal Practice and of Psychotherapy Research - 284 . , 6, 276 . A , Shaw Beck . A. T., Rush : Emery ,J , B. F., & , G. (1987 ). Cognitive . New therapy of depression York : GuilfordPress . : Basics Beck . J. S. (1995 andbeyond ). Cognitive . NewYork: GuilfordPress . therapy Black . I., Adler , J., Dreyfuss , C., Friedman ., & : Roach , W., LaGamrna ,E , A. (1988 and ). Experience the biochemistry of informationstoragein the nervoussystem . In M. Gazzaniga .), Perspectives (Ed in memory research . (pp. 3- 22 ). Cambridge , MA: MIT Press : Grossberg . (1987 , G. A., & architecture ). A massively for a self Carpenter - organizing ,S parallel neural machine . Vision , Graphics , andImage pattern . recognition Computer , 37 , 54- 115 Processing ., Bloom : Roth ,J , F., & , R. (1991 basis ). Thebiochemical Cooper , 6th ed. New ofneurophannacology York : Cambridge Press . University Corwin ., Fennan , J., Peselow ,E : Fieve , K., Rotrosen , J., & , R. (1990 of decision in ). Disorders - 833 affective disease : An effectof p-adrenergic ? Biological . , 27 dysfunction , 813 Psychiatry Deakin , J. F. W. (1994 ). Theclinicalanxiolyticandantidepressant of drugswith actions efficacy on serotonin . In J. DenBoer& : J. M. AdSitsen .), Handbook systems andan (Eds .riely of depression - 472 (pp. 447 . ). NewYork: MarcelDekker

160

BehavioralStates

in prefrontal cortexof affective . , Robbins : Roberts Dias , A. C. (1996 ,R , T. W., & ). Dissodation . . Nature andattentional shifts , 380 , 69- 72 andneuropsychologic of : Geder Es , L. (1996 , P. J., Grattan , L. M., & ). Neurologic aspeds linger .), Head . In M. Rizzo& : D. Tranel(Eds in postconcussive frontal lobe impainnents syndrome - 440 : Church ill Uvingstone . ). New York (pp. 415 syndrome injuryandpostconcussive Press . . NewYork : Cambridge Fauconnier , G. (1994 ). Mentalspaces University . : Cambridge neurons . Cambridge Fillenz , M. (1990 ). Noradrenergic UniversityPress . state . Hillsdale : Erlbaurn : frompsychology to brain Gilbert , NJ , P. (1984 ). Depression . . NewYork : Cambridge , J. A. (1987 ). The UniversityPress offearandstress Gray psychology Review code ? Psychological , 87 , 1 51. ). How doesa brainbuilda cognitive Grossberg , S. (1980 mind . London : Sage . discursive Harre ., & : Gillett , G. (1994 ,R ). The . andrecovery . New York : Basic Books Herman , J. L. (1992 ). Trauma - depressive illness . In D. S. Levine & : S. J. networktheoryof manic Hestenes , D. (1992 ). A neural - 257 in neural networks Leven(Eds .), Motivation , , emotion , andgoaldirection ). Hillsdale (pp. 209 : Erlbaurn . NJ : A model . Science with neural drcuits : Tank , 625 , 233 , J. J., & . D. W. (1986 ). Computing Hopfield . 633 - 183 . Science . of dedsions : Tversky , 250 , 181 , A. (1982 Kahneman . D., & ). Theframing on . Presented at the Conference .AiM.: A triune extension to theART model Leven , S. J. (1987 ). S . Architecture 9, 1987 in BrainandComputer Networks , Denton , TX, October at the M.I.N.D. : A triune . Presented andlearned Leven , S. J. (1988 ). Memory approach helplessness . Dallas on Motivation , May 23, Conference , andGoalDirectionin NeuralNetworks , Emotion . 1988 . In D. S. Levine& : of hope Leven , andthe dynamics , memory , S. J. (1992 ). Learned helplessness - 300 in neuralnetworks .), Motivation , emotion , and goal direction (pp. 259 ). J. S. Leven(Eds : Erlbaurn . Hillsdale , NJ . Presented at the American as a dynamical Leven , S. J. (1993 ). Depression system Psychological 10 . 1993 Annual , , June Sodety Meeting , Chicago neuralnet user : Toolsfor effective , creative , andneurotic Leven , S. J. (1995 ). Intelligent agents - 304 . : Erlbaurn 2 . . Mahwah Neural Networks Vol. 293 . In World on , , , ) NJ ( Congress pp support underuncertainty . embedded networks . J., & : Elsberry , W. R. (1990 Leven ,S ). Interactions among - 742 : on Neural Networks In I]CNN International , Vol. 3, (pp. 739 ). SanDiego JointConference IEEE . anddysfunction . .). (1991 : Benton ., Eisenberg Levin , A L. (Eds ). frontallobe , H. S function , H. M., & . Press : Oxford University NewYork : Erlbaurn . . Hillsdale to neural andcognitive . (1991 Levine , NJ , D. S ). Introduction modeling in .). (1992 . In Motivation . (Eds Levine : Leven , emotion , andgoaldirection , S. J , D. S., & ). Preface . : Erlbaurn neural networks . Hillsdale , NJ ., Leven : Prueitt Levine , and the frontal , disintegration , P. S. (1992 , D. S , S. J., & ). Integration in neural networks .), Motivation & : S. J. Leven lobes . In D. S . Levine , and , emotion (Eds goaldirection - 335 : Erlbaurn . , NJ (pp. 301 ). Hillsdale in issues and theoretical : Prueitt Levine , P. S. (1993 , D. S., Parks , R. W., & ). Methodological . International , 72 , neuralnetworkmodelsof frontal cognitivefunctions Journal of Neuroscience . 209 - 233

and Depression Leven : LearnedHelplessness

Lowen, A . (1971). Thel Imguage . New York: Collier Books . of the body models of hysteresis . IEEETransactions on Magnetics Mayergoyz, I. D. (1989). Dynamic Preisach , 24, 2925- 2927. Maclean , P. D. (1990). The triune brain in evolution : Role in paleocerebral . New York: functions PlenumPress . Miller , G., Galanter : Pribram , E., & , K. (1960 . New York: ). Plansand the structureof behavior RandomHouse . Minor , T. R., Dess : Overmier, J. B. (1991). Inverting the traditional view of ' 1earned , N. K., & " . In MR .), Fear , avoidance , and phobias(pp. 87- 133). Hillsdale helplessness Denny (Ed , NJ: Erlbaum . -Verlag. , R. (1985). Chemoarchitecture Nieuwenhuys of the brain. New York: Springer ' : Greenberger , C. A ., & , D. (1995). Clinicians guide to mind over mood . New York: Padesky Guilford Publications , Inc.

Peterson ., & : Seligman , C., Maler ,S , M. (1993 . New York : Oxford University ). Learned helplessness Press . : Hendrickse , F., Kramer , G., & , W. (1993 . In J. Mann& Petty : D. Kupfer ). GABA anddepression .), Biology disorders (Eds : Plenum , PartA (pp. 79- 108 Press . of depressive ). NewYork Pribram . (1991 andperception . Hillsdale ,K : Erlbaum ). Brain . , NJ ' Pribram : Gill, M. (1976 s project , K., & reassessed . New York : Basic Books . ). Freud Pribram : McGuinness , K., & , D. (1975 , activation ). Arousal . , andeffortin the controlof attention - 149 Review . , 82 , 116 Psychological Richelson , E. (1997 of newantidepressants ) . Pharrnacokinetic : A reviewof the drug interactions - 847 effects on themetabolism of otherdrugs . MayoClinical . , 72 , 835 Proceedings Rumelhart , D. E., Smolensky t Hinton , P., McOelland . (1986 , j . L., & andsequential , GiE ). Sdtemata es in POPmodels . In j . L. McClelland & tO . Rumelhart .), Parallel distributed thoughtprocess (Eds , Vol. 2 (p~. 7- 57 . ). Cambridge , MA: MIT Press processing . A , Mirin, S. M., Hauser Samson ,J t Schildkraut , S. T., Fenton . (1992 , B. T., & , J-J ). Learned help lessness and urinaryMHPG levelsin unipolardepression . American , 149 , Journal of Psychiatry 806 - 809 . . , Orsulak , J., Bloomingdale ,K . P., Gerson t , B., Kizuka . P., Cole Schatzberg , A., Samson , J., & Schildkraut , J. (1989 classmcation of depressive ). Towarda biochemical disorders , X: Urinary catecholamines , their metabolites in subgroups , and O-type scores of depressive disorders . Archives - 268 . , 46 , 260 of General Psychiatry Schildkraut , J. (1965 of affective ). Thecatecholamine disorders : a reviewof supporting hypothesis - 522 evidence . American . , 122 , 509 Journal of Psychiatry Schneider . E., Alavi, A., Seligman , F., Cur, R . J. Mozley , M. E., Mozley , L. H., Smith : ,R , P. D., & Cur, R . C. (1996 bloodflow changes in limbicregions ). Cerebral induced by unsolvable anagram tasks . American - 212 . , 153 , 206 Journal of Psychiatry Self , D. W., Barnhart : Nestler , W. J., Lehman , D. A , & , D. (1996 of ). Oppositemodulation -seeking cocaine behavior . Science by Dl - and D2-like dopamine , 271 , 1586 receptor agonists . 1589 . NewYork , M. E. (1975 : W. H. Freeman . ). Helplessness Seligman Suomi . J. (1991 ,S models of affective ). Primate disorders . In J. Madden .), Neu , IV (Ed separation emotion and . 214 . , , 195 New York : Raven Press . ( robiology of learning ) affectpp

162

BehavioralStates

: Steinberg Ialdisorder. New York: Wiley . Tyrer, P., & , D. (1993). Modelsfor mtn Van der Kolk, 8. A . ( Ed .). (1981). Psychological trauma . Washington, DC: American Psydtiatric . Press Welss -induced depression : Critical neurochemicaland electrophysiological , J. M . (1991). Stress . In . IV . Madden , , and affect(pp. 123- 154). New J ( Ed ), Neurobiology of ltaming, motion changes York: Raven .

Welss : Simson of the locuscoeruleus : Implications forstress , J., & , P. (1989 ). Electrophysiology induceddepression . In G. Koob : D. Kupfer(Eds .), Animal models , C. Ehlers ,& of depression - 134 : Birkhiuser . (pp. 111 ). Boston Werbos : New toolsfor prediction andanalysis in thebehavioral , P. J. (1974 ). Beyond regression . Unpublished sciences PhiD . dissertation . Reprinted in , HarvardUniversity , Cambridge , MA Werbos : From Hon ordered deriva Hves to neural networks and , P. J. (1993 ). Theroots of bR Ckpropaga . NewYork : Wiley. Hcalforecas Hng Poll : A psychobiological Willner . New York : Wiley-interscience . , P. (1985 ). Depression synthesis

163

Leven : LearnedHelplessness and Depression

Waking and Sleeping States JeffreyP. SuttonandJ. Allan Hobson

Waking and sleepare behavioral statespresentin all mammalsand the rhyth , Bergmann , micity between thesestatesis essentialto survival (Rechtschaffen Everson , et al., 1989). Sleep changesduring development and aging (Roffwarg , Muzio , and Dement, 1966), and it subserves important functions related to energy conservation control, appetite , neuroimmunoendocrine , and thermoregulation (Moore , 1979). In addition, sleep is linked to memory, learning, and other cognitive functions (Hobson, 1988b), and there is a longstanding , although largely inferential, notion that sleep is important for affective regulation (Kramer & Roth, 1972). There are also well-established correlationsbetween disturbances in sleepand disordersof brain and mind. From the perspective of modeling, studies of sleep fall into three main . The first category concerns circadian and other biological categories e. rhythms ( g ., ultradian and infradian rhythms). Severalmathematicalmodels have been proposed that capture features of these rhythms, as well as -wake and seasonalcycles the coupled oscillations among endocrine , sleep , Czeisler , Pilato, et al., 1982, Strogatz , 1986, Moore -Ede& Czeisler (Kronauer , 1984). These chronobiological models do not necessarilyincorporate ana tomical or neurophysiologicaldetails of sleep . A secondcategory of models dealswith circadiancontrol and modulation on detailed neurobiological systems . While some integrative work regarding the molecularbiology and geneticsof circadianrhythms has been performed , 1995), most investigations in this category have focusedon individual ( Takahashi neurons and small networks. Models of cellular networks have been helpful in synthesizing information about brainstem circuitry and thalamo cortical interactions(Steriade , Pare , Parent , et al., 1993; McCarley & Hobson, 1975). Furthermore have been confirmed by induced , many predictions -wake physiology (Steriade behavioral or pharmacologicalalterationsin sleep & McCarley, 1990). The third category of models uses large-scaleneural network models to examinethe effectsof sleepon cognitive process es. Models in this category have lessanatomicaldetail than the models in the secondcategory, but they emphasizethe cooperative effects among many neuron-like units. They are useful for exploring some of the computational principles that may underlie

information storage , transfer , sequencing, and learning during sleep (Sutton , : Hobson , et al., 1992a). Generally , this is accomplished by Mamelak, & network studying properties during simulated states of neuromodulation that characterize wake and sleep. In this chapter , we present an overview of neural network models of sleep. We briefly review some of the relevant behavioral , cognitive , and physio logical data, and then describe different models that attempt to integrate these data. Our emphasis is on how shifts in the relative balances of aminer gic and cholinergic modulation , occurring during waking and sleep, influence information sequencing and learning in large - scale networks , including the cerebral cortex . We suggest that modeling the effects of neuromodulation on network dynamics is a useful way to study some aspects of sleep. Moreover , we claim that similar approach es are relevant to modeling a variety of behavioral states and disorders affecting the brain and cognition .

At CONSIDERA EXPERIMENT nO NS
General Features

-wake cycle consists of three main behavioral states The sleep : the waking state , nonrapid eye movement ( NREM , and rapid eye movement ) sleep . Each state has distinctive behavioral and electrophysiological (REM) sleep as well as characteristicneurophysiological features . In humans , , properties different cognitive attributes have been ascribedto each of the three states (figure 7.1).

'.1 Chart sUJI behavioral manifestations , electrophysiological , and psychological Figure 7 Unarizing of the wakingstateand NREM and REM sleepin humans . EMG , electromyelogram ; EEG - oculogram . ( Modified fromHobson , electroencephalogram & : Steriade .) ; EOG , electro , 1986

166

BehavioralStates

. These , skeletaland eye movementsare abundant During the waking state diminish significantly during NREM sleep , which accountsfor 80 percent of all sleep in adult humans . Approximately six episodesof REM sleep , each , are interspersedamong the NREM sleep lasting eight to fifteen minutes periods. REM sleepis partially distinguishedby skeletalmuscleatonia, along with phasicbursting eye movements(Hobson & : Steriade , 1986). The waking state and REM sleepare correlated with low -amplitude, high: Kleitman, (EEG ) recordings (Aserinsky & frequency electroencephalogram 1953). In contrast , NREM sleep is manifest by high-amplitude, lowfrequencywave . To completely specify state , sleep laboratory paradigms involve of behavioral information , and , the EEG recordings typically other physiological recordings the electro EGG , including ), elec oculogram ( , and heart rate (Hobson, 1988a ; tromyelogram (EMG), core temperature Somers , Dyken; Mark, et al., 1993).
Cognitive Properties

-wake cycle, there are changesin mental process es that Throughout the sleep alter perception , thought form and content, orientation, and mood (Bootzin, Kihlstrom, & Schacter , 1990). Dreaming is the hallmark psychological experience . It is known to be predominantly a REM sleep characterizingsleep phenomenon (Aserinsky & Kleitman, 1953; Dement & Kleitman, 1957). Dreams have been described since biblical times, and despite voluminous anecdotal reports and intuitions about dreaming (Brook, 1987), relatively few rigorous scientific studies have been made of this fascinating state of mind (Hobson, 1988a , of course , is that it is difficult ). A fundamentalobstacle to elucidate exact relationshipsbetween sleep and cognition, since observations are limited to psychophysical data collected during behavioral states that border on sleep(i.e., the waking state , falling asleep , waking up). Nevertheless , there is growing scientific evidence that cognition is state es are associated , in the sensethat different types of cognitive process dependent with different neurophysiological states . These states include, but are not limited to sleep statesof anxiety, meditation, and drug, and encompass . In the waking state induced psychosis , thoughts are generally logical, . In NREM sleep consistent , and continuous , thought processing typically consistsof ruminative sequences (Antrobus, 1983), not unlike the perseveration -compulsive disorder. In contrast or rumination that occurs in obsessive to the waking state and NREM sleep , REM sleep is associatedwith graphic and formally bizarre images containing discontinuities , incongruities, and uncertainties (McCarley & Hoffman, 1981; Foulkes , 1985). The predominantly external perceptualcuesof the waking state shift to internal sources . during sleepand dreaming In addition to changesin thinking associatedwith different states , there . These are usually described are alterations in other aspectsof the mental status qualitatively, with considerablefocus on the content and inferred

167

Sutton & : Hobson: Waking and SleepingStates

, 1900). However, some excellent quantitative meaning of dreams (Freud work has also been conducted on the formal properties of cognition during , 1994). For example , Hobson and co-workers have sleep (Baars 8t Banks how , sensory perception differs in waking and sleep ( Williams quantified Merritt , Ritlenhouse , et al., 1992). -dependent is One impressiveway in which cognition appearsto be state with respect to discontinuities in visual attention and orientation (Hobson, Hoffman, Helford, et al., 1987). In sleep , and specificallyREM sleep , shifts in attention during dreamsarereported to occurmore frequently relative to waking . The shifts are often suddenand unexpected . They are also experiencedin associationwith intenseemotion, especiallyanxiety (Merritt , Stickgold, Pace Schott, et al., 1994; Sutton, Ritlenhouse Pace schott al. et . , , , 1994b) The magnitude of discontinuities in attention can be determined from narrative reports describing experiencesduring the waking state and sleep . Narrative graphing is a reliable techniquedeveloped for mapping language -Schott, et al., basedreports onto geometric graphs (Sutton, Ritlenhouse , Pace . 1994a The method is not specific for state ) , and the graphs are helpful in , which are determining the magnitude of discontinuities in visual scenes reflective of changesin attention (figure 7.2a). They are also useful for representing of information and thought (figure 7.2b). Narrative temporal sequences has been used to map, in detail, discontinuities and bifurcations graphing in information sequencingin dream reports (Sutton et al., 1994a ). In addition to changesin thought sequencing , memory and learning have -dependent features state . Memory and learning are intermingled process es. with as well as with state , with maximal learning They vary development occurring early in ontogeny and decreasingwith age. This pattern is loosely correlatedwith the amount of REM sleep , which can reachlevels of 50 to 80 in life et al. 1966 , ( ). There is also an interesting relationship percent early Roffwarg between the integrity of REM sleep and the ability to learn visual attention tasks (Karni, Tanne , Rubeastein , et al., 1994). It is known that , Kactaniotis , 8t learning is impaired by REM sleep deprivation (Fishbein Chattman in REM sleepfollow periods of learning , 1974), and that increases in adult humans(Smith, Kitahama , Valatx, et al., 1980). With new techniques of multiunit recordings and noninvasive neuroimaging , such as functional magnetic resonance imaging (MRI ), it is becoming possible to examine realtime brain changes associatedwith learning (e.g ., Recanzone , Merzenich, , et al., 1992; Karni Meyer, Vezzard Jenkins , et al., 1995). It will be very -dependent aspectsof learning with these interesting to examine the state new technologies . Some progress has already been made in this direction ( Wilson8t Mc Naughton, 1994).
Neurophysiological Data

At the cellular level, the biogenic aminesare chemicalsubstances known to be important for memory, attention, and learning (Flicker : , McCarley, &

168

BehavioralStates

8 7 >~ c .g u 6 5 . 4 3 \ . "

8 time

10

11

12

Figure 7.2 a, Hierarchy graph representinga narrative report. Starting with a written account of visual imagery (not shown), objects are identified in the report and mappedonto a graph following a set of reliable rules (Sutton et aI., 1994a ). Different levels of the hierarchy are determined basedon part-whole relationships . The narrative segmentitself is the zeroth level, or root node, of the hierarchy . Two principal scenes are identified in the examplegraph. One sceneis a houseand the secondsceneis a bedroom (in a different housewith some friends). These scenes form first-level objects . Parentsand a painting are contained within the house and make up second -level objects . At the third level, two parentsare identified along with three objects namedin the painting. b, Once the hierarchy graph is obtained , the time course of attention among the visual objects can be plotted. Eachtransition between two objects is assigneda measureof discontinuity based on how far the transition moves within the hierarchy graph (Sutton et al., 1994a ). The extent of discontinuity is plotted as a function of time, as shown in (b). It has been reported that large discontinuities (e.g ., at time step 12) are correlated with subjective experiences of intenseemotion (Sutton et al., 1994b .) ). (Modified from Sutton et aI., 1994a

169

Sutton & : Hobson: Waking and SleepingStates

friends /

"

"

frame

"

bedroom

segment

water

painting

"

'

narrative

ships

house

mother

parents

father

level

13

14

15

Hobson, 1981; Schmajuk& Moore, 1985; Jarbe , Callenholm , Mahammed , et al., 1981 ; Kandel , Klein, Hochner , et al., 1987; Hopkins & Johnston , 1988). It is thereforenot surprising to discover that changesin the relative balances of some of these chemicals are associatedwith different waking and sleep states . The details of the changeshelp to explain some well-known phenomena , suchas the amnesiaoccurring with sleep(Hobson, 1988a , ). During sleep the main chemical changesinvolve norepinephrine (NE), serotonin (5-HT ), and acetylcholine(A Ch). Much of the knowledge about brain chemistry during sleep has been gleaned from studies of behavioral state control at the brainstern level (Steriade & McCarley, 1990; Datta, 1995). There has also been excellent work done to elucidate the activity of forebrain neural populations during the waking state and sleep (Wilson & Mc Naughton, 1993; Poe , Rector, & . 1994 Within the brain stem cells which are most active , pontine ) Harper, , release the monoamines during the waking state and silent during REM sleep NE and 5-HT from the locus ceruleus and dorsal raphe nucleus , Hobson & . the 1975 cell , , ( , ) McCarley Wyzirki many respectively Among populations that becomeactive during REM sleep are cells associatedwith A Ch release . These cells , located in the peribrachial pons, are generally , except during the startle responsewhen they fire briefly but dramatically quiescent , the cessation (Bowker & Morris on, 1976). In NREM and REM sleep of NE and 5-HT neural activation is coupled with relative increases in cholinergic activity . Some of the principal pathways are summarizedin figure 7.3. While the aminergicand cholinergic systemsaccountfor lessthan approximately 5 percent of all neurons , they are strategically positioned to induce effects on cortical circuits. Both NE and 5-HT widespreadneuromodulatory inhibit the spontaneousfiring rate of certain forebrain neurons (Aghajanian & Van der Maelen, 1986), and they increase the signal-to -noise ratio in -Schreiber neural networks (Segal , 1985 ; Servan , Printz, & Cohen, 1990). In NREM sleep the forebrain receives , approximately 50 percent lessaminergic modulation compared to the waking state. In REM sleep , the forebrain is , 1986). These relatively free of aminergic modulation (Hobson & Steriade in result circuits with increased changes widespread spontaneous neural activation. , and Cholinoceptive neurons exist throughout the cortex, hippocampus thalamus (Olivier , Parent , & Poirer, 1970; Bear & Singer , 1986). Primary sourcesof A Ch to the cortex are from neurons in the basal forebrain, while the thalamusreceivesits supply from peribrachial neuron projections (Steriade et aI., 1988). The principle action of A Ch releaseis thought to facilitate neural firing (Steriade& ilinas , 1988). Global cholinergic activity in NREM sleepdeclinesto approximately 50 percent of its value found in the waking state ; but it increasesto a level comparableto that present in the waking state during REM sleep . This is due to the effect of phasic bursting from brainstem neurons which generatecharacteristicEEG tracings termed ponto

170

BehavioralStates

cortex
Sf

thalamus

brainstem
VI , I IV , ~ III Ch A 15HT NE

spinal

cord
. . . Inhibitory ~ Excitatory

' .3 Schematicnetwork of waking and sleep regulation. Four levels of brain organization Figure 7 are shown with representativeexcitatory (open neurons ) and inhibitory (solid neurons ) . Reciprocally connected aminergic and cholinergic systems within the brain stem pathways induce widespread modulatory effects . Aminergic disinhibition and cholinergic excitation in REM sleep are associatedwith rapid eye movements , EEG desynchronization , and ponto- occipital burst cell (pGO) waves . Muscle twitches occur when phasic excitation overcomes geniculo tonic inhibition within the spinal cord. Sutton , Mamelak , and Hobson (1992b) have simulated the simplified cortical architecture(heavy outline) on a Sun workstation . cr , cortical cell; ST, stellatecell; PT, pyramidal cen ; TC, thalamocorticalcell; NE, norepinephrine ; SHT, sero tonin; Ach, acetylcholine ; 1lI, oculomotor, IV , trochlear , and V , trigeminal motor nuclei. (Adapted from Hobson , 199Ob , and Sutton and Hobson , 1994.)

- occipital (PGO ) waves(Brooks & Bizzi, 1963; Steriade& Uinas, 1988). geniculo This cholinergically mediated stimulation conveys information about the direction of eye movements which become , in REM sleep , completely from external stimulus control , Lydic, Baghdoyan (Callaway , et decoupled al., 1987). In humans , evidence for PGO activity has been demonstrated by Winkelman , , and Duffy (1983), using EEG mapping techniques . McCarley Functional neuroimaging using positron emission tomography (PET ) has demonstrated decreased glucose utilization in the hippocampus during

Sutton & : Hobson: Waking and SleepingStates

NREM sleep relative to waking and REM sleep (Ramm & : Frost, 1986; Dive Salmon al. , , et , 1990). Decreasedmetabolism in the thalamus Maquet, and midbrain has also been observed during sleep deprivation (Wu, Gillin , Buchsbaum , et al., 1991). These findings support the notion that human like , , is a hypoadrenergic , hypercholinergic sleep sleep in other mammals state relative to the waking state .
Sleep and Neuropsychiatric Illness

There are several brain and sleep disorders that are characterized by abnor malities of the aminergic and cholinergic systems (Hobson & : Steriade, 1986 ). Clinical depression , for instance, is a disorder associated with altered catecholamine balance. An excellent biological marker for depression is a sleep disturbance , which is partially manifest by shortening of the latency to the first REM sleep period (Gillin & : Borbely , 1985 ). The cognitive impairments and pharmacological treatments of this common mood disorder are consistent with it being a hypoadrenergic , hypercholinergic state. The reader is referred to Hobson and Steriade ( 1986 ) and Montplaisir and Godbout ( 1990 ) for further discussion of the relationships between sleep and disorders in psychiatry and medicine .

COMPUTATIONAL MODELS
Overview

There is a long history of using neural network models to study states of brain and mind related to waking and sleep . Freud , for example , developed a model of neuronslinked by artificial synapses into a network driven by nones (Freud , 1895). physiological energy forces that altered consciousprocess McCulloch, the psychiatrist and modeler known, in part, for his seminal work with Pitts on networks of binary valued neurons , examined network models of altered state process es , including delusions (McCulloch & Pitts, 1943; McCulloch, 1965). These and other models laid the groundwork for more recent theoretical work that specifically address es neurophysiologically based correlates of and . As outlined earlier , there are two broad, yet partially waking states sleep . One category deals with state overlapping, categories of models control, and focusesprimarily on brainstem and thalamic systems , while the secondcategory examinesforebrain networks and the effects of modulation on large-scale neural populations. In common with all brain models , simplifications of anatomy and physiology are required to make investigations tractable . One of the great challengesin neural modeling is determining how to reduce the complexity of biological systems to foster rigorous study while maintaining pertinent details. Examplesof such simplifications are discussed in this section with regard to computational modeling.

BehavioralStates

Brainstem and Thalamic Models

The phenomenology of REM sleep as a rhythm mediated by neural networks has been investigated in several complementary ways. In 1975, McCarley and Hobson proposed a mathematicalmodel of reciprocally interacting : Hobson, 1975). As populations of brainstem neurons (McCarley 81 illustrated in figure 7.4a, one population correspondedto aminergic cells that inhibited a second population of neurons , as well as itself. The second , with excitatory projections to the population was cholinergically mediated aminergic cells and back to the cholinergically mediated population. Using Volterra-Lotka equationsemployed in simple field biology models of preda tors and prey (Volterra, 1931; Lotka, 1956), the reciprocal interaction model displayed oscillatory behavior consistent with physiological behavior regulating -wake cycles (McCarley 81 : Hobson, 1975). The model also made sleep several predictions regarding aminergic and cholinergic interactions , and of these have been confirmed Hobson 81 : Steriade , 1986). ( many subsequently One of the limitations of the reciprocal interaction model was that it was exquisitely sensitive to initial conditions. It would have to be reset each , a limit -cycle model was night . To combat this and other shortcomings introduced that took into considerationmore physiological details, including constraints on neuron firing rates (McCarley 81 : Massaquoi , 1986). Other investigators usedmodelsof brainstemand thalamiccircuitry to demonstrate how transitions in physiological state may be associated with changesin the of individual neurons and neural , McCor dynamics populations (Steriade mick, 81 : Sejnowski , 1993). In particular, thalamocorticaloscillations and resonance among neural populations have been a topic of considerablerecent interest. In one approach , oscillations critical to binding neural populations are modulated in sleep and waking by different sensory inputs (ilinas & : . 1994 ) Ribary,
Distributed and Cortical Network Models

In contrast to the models discussed in the previous subsection , large - scale models sacrifice relatively greater anatomical and physiological detail in exchange for a better understanding of computational principles involving parallel distributed networks . Large- scale models typically address the following question : What are the functions associated with different states of correlated physiological and mental activity ? From a computational , as well as a biological perspective , different states of operation presumably have specialized tasks that allow the system as a whole to perform in an adaptive , diverse , and efficient manner. A common theme of large - scale models is that changes in modulation specify state and provide clues to the types of information processing that occur. Surprisingly , very little computational work has actually been done on sleep and waking using such networks .

Sutton& t Hobson : WakingandSleeping States

t~

v
1 -

PIO

BehavioralStates

tit "
. 9 9 9

9 V
~ 1IO JV
~

8 Ploq

.tU

Ap A -

atf uaa a ) '

~ ~

Sutton & : Hobson: Waking and SleepingStates

tq

-dependent processing based on neuro The first model to detail state physiological changesoccurring during waking and sleepwas the activationsynthesis hypothesis of Hobson and McCarley (1977). In their model, activation was due to PGO bursting. This originated in the pontine brain stem and drove the cortex in REM sleep to generate , or at least facilitate, the cognitive processof dreaming . The model posited a brain-basednotion of dreaming that departed from previous psychological theories of dreaming (e.g ., Freud , 1900). The importance of phasic PGO bursting during REM sleep has also be . PGO activity is related to theta rhythms, recognized by other researchers which have been proposed to mediate memory encoding during REM sleep , 1990). This is an interesting suggestion given the role of phasic ( Winson A Ch in some models of memory. One mechanismof action may be that A Ch excites inhibitory neurons , which in turn hyperpolarize or clamp other neurons during the encoding of novel inputs (Hasslemo , 1993). The PGO system may also be relevant to memory consolidation according to a two stage hippocampal learning model proposed by Buzsaki (1989). Off -line consolidation during NREM sleep is also a central feature of interleaved -cortical model recently learning in a hippocampal proposed by McClelland, ' and O . 1994 , ( ) McNavghton Reilly In general , neuromodulatory changesin sleep have been used metaphori: cally in a variety of neural network models of associativememory (Geszti & pazrnandi , 1987). Phasetransitions in the dynamicsof distributed neural networks have also been attributed to inhibition associated with 5-HT levels in simulatedwaking and sleepstates(Nakao, Takahashi Mizutani , , et al., 1990). Sometimes in the , striking changes phaseor state of an artificial neural network are referencedin the context of sleep and waking states , to emphasize -wake physishifts in processingrather than to make a connection with sleep Hinton et . 1995 , ( , , al., ) Dayan Frey ology A provocative hypothesis postulated by Crick and Mitchison (1983) suggestedthat PGO activation during REM sleep may have an unlearning effect to help prevent memory overload. Their argument , which was independentl Feinstein and Palmer 1983 , , ( ) and Clark, proposed by Hopfield Rafelski , and Winston (1985) in different contexts was basedon the utility , of random PGO-like inputs for unlearning within nonsequencing neural networks operating near maximal capacity . The universality of unlearning as a computational principle has gained increasedattention in recent years : van Hemmen , Klemmer ( Wimbauer ,& , 1994). However, it is not known how PGO-like inputs might affect networks with temporal sequencingfeaturesor what roles aminergicwithdrawal may play in unlearning . Mamelak and Hobson (1989) incorporated PGO phasic activation with arninergicwithdrawal in a model of forebrain demodulation during sleep(see : figure 7.4). Their model built on the activation-synthesismodel (Hobson & McCarley, 1977). It suggested that bifurcations and bizarrenessin dreams

176

BehavioralStates

may be due to combined aminergic and cholinergic interactions . The aminer gic system in the model served as a random noise filter or gain function (Segal, 1985 ; Keeler , Pichler , & Ross, 1989 ). Aminergic withdrawal was tonic or steady , and it resulted in a flattening of the sigmoid response curve that relates the probability of a neuron firing to its firing threshold (Little , 1974 ; Freeman, 1975 ). This relationship , which is one of several effects attributed to the biogenic amines, is shown in figure 7.4c. The facilitation induced by phasic cholinergic excitation is illustrated in figure 7.4d . The combined effects of these modulatory changes on sleep and other states has also been discussed in depth by Hobson ( 1990 ). Sutton et al. ( 1992b ) tested the model of Mamelak and co- worker through computer simulations on a temporal ~equencing network . They confirmed that bifurcations between embedded sequences were indeed possible with changes in simulated aminergic and cholinergic modulation . Perhaps more impressive was the rich repertoire of sequencing and learning possibilities associated with changes in modulation in a state of simulated REM sleep. Apparently random alterations , such as increased synaptic noise due to aminergic withdrawal and intermittent global excitation due to PGO bursting , did not yield random outputs . Rather , the network remained highly constrained in its information capabilities . Seemingly bizarre outputs had a deep structure linked to the embedded sequencing and learning rules. While the exact relationships between the simulation results and neurobiology remain unclear, the emergence of deep order in the context of demodulation , or noise, is now a well -recognized phenomenon (Collins , Chow , & Imhoff , 1995 ; Moss & Wiesenfeld , 1995 ). This is discussed further in the next two sections.

-DEPENDENT STATE AND MODULA nON SEQUENCING


General Considerations

Distributed network models that attempt to link psychological events with underlying neural dynamics in waking and sleep rely primarily on the idea that physiological state changesare regulated by changesin neuromodulation -wake physiology, there is an impressive . While this conceptappliesto sleep literature demonstrating a spectrum of behaviors in neural networks , 1985; Davis, Jacobs subjected to changesin neuromodulation (Selverston , & : Schoenfield : Marder, 1991 , 1989 ; Harris-Warrick & ; Silberstein , 1995). Furthermore , the notion of critically altering the dynamic behavior of a system , applies not only by varying a quantity, suchas a neuromodulatory chemical to neural systemsbut to many systemsin biology and physics . Indeed , there is an analogy between transitions in physiological states and transitions in physical states(Sutton, 1995).

177

Sutton at Hobson: Waking and SleepingStates

Assumptions

in

Modeling

There are many different ways to study distributed networks and to investigate the effects of modulation and state transitions (Anderson and Rose nfeld, 1988; Hobson, Mamelak : Sutton, 1992). In bridging aspects of , & neuromodulation with changes in state - dependent cognition, and discontinuities in visual imagery and memory in particular, a reasonableapproach has been to examine networks with two essential features . The networks should generate temporal sequences of cognitive information and the networks should simulate the effects of aminergic and cholinergic modulation. We have incorporated these criteria into an impressionistic model that is outlined here (Sutton et al, 1992b ; seefigure 7.3). We maintain that the rich behaviors associatedwith our simplistic model underscore the enormous - dependentbehaviors present in mammalian systems complexities of state and . The simplicity attempts to extract key conceptsat during waking sleep the expenseof details , and we emphasizethat the modeling should not be construed as an effort to replicate the behavior of the cerebral cortex and . deeperstructuresin waking and sleep To make explicit what we have attempted to do, we list the pertinent assumptionsand limitations below: 1. The network consistsof model neurons , and the collective activity of the neurons at a given instant in time generates a pattern of neuron firing throughout the network. -stable 2. A small fraction of the firing patterns are quasi . Once the network enters a quasi stable pattern, it remains in that pattern for a period of time -stable patterns serve as the netbefore moving to another pattern. The quasi ' work s memory. -stable 3. The ability to store information about the memory, or quasi , patterns is encoded distributively among the connections linking individual neuronswithin the network. 4. Information about how particular memory patterns are temporally linked to eachother is also distributively stored among the connections . When the network enters one memory pattern, it sequentiallymakesa transition to the next stored pattern in the sequence . 5. The dynamics of the network are governed by a computational rule that . Each neuron , parallel manner updates all of the neurons in a synchronous determinesits inputs at a given instant in time, and after comparing the sum of the inputs with its threshold value, each neuron determines whether or not it fires. Once a neuron is firing , it requires continued input to remain . The updating rules on firing and the , it becomesquiescent firing ; otherwise evolution of the network are temporal partially based on physiology and based on convenience , Gat, et al., (Abeles, Bergman partially computational 1995).

178

BehavioralStates

6. Inputs from the aminergic and cholinergic systems are incorporated at eachtime step in the evolution of the network. 7. The instantaneous output of the network is given by the global pattern of . This can be expressedas an overlap between activity among the neurons the instantaneous pattern and eachof the embeddedmemory patterns. Complete overlap with a single memory pattern signifies that the network is in that memory. The network may partially overlap with a combination of , or it may be in a random pattern with no obvious overlap memory patterns with any memory pattern.

Neural Network Model


Many attractor neural networks can be modified to have the required sequencingproperties (Amit , 1989; Reiss & Taylor , 1991). We chose to adapt a model which had well-characterizedtemporal sequencingproperties and which was amenable to adding neuromodulatory features (Kleinfeld, 1986; Sompolinsky & Kanter, 1986; Gutfreund & Mezard, 1988). In our network of memories , and sequences , memory patterns encoded visual memories visual or flows among images (Sutton representedtemporal orderings the case where two . For illustrative we considered et al., 1992b , ) purposes of memory patterns were stored in the network. Both sequences sequences contained three memory patterns each , and these were arranged into two . A consisted of memory patterns AI , Al , and A3 , which made loops Loop transitions sequentiallyas Al - + Al - + A3 - + Al . . . . Similarly, loop B comprised the sequenceof memory patterns B1 - + B2 - + B3 - + B1 . . . . Continuities and discontinuities of visual information were representedby the to degree which transitions among memory patterns in the model remained within or jumped between loops A and B. Aminergic regulation was modeled by intrinsic noise at the level of the (seefigure 7.4c). We consolidatedNE and 5-HT together, although synapses it was recognizedthat thesemodulators have subtypesand heterogeneity of function. Cholinergic facilitation was an extrinsically generatedinput which nonselectively excited all the neurons in the network (see figure 7.4d). The connection weights among the neurons and the membranethresholds were held constant during the simulations , although the effects of learning were . A mathematical outline of the model is provided in examined subsequently interested reader is referred to Sutton et al. (1992b) for the appendix and the , details.

SIMULA ANDLEARNING nON RFSULTS


Simulation Results

The modelsparameters were selectedto allow the network to generate , in .A and stable of memories the simulated waking state , repetitive sequences

179

: Waking andSleeping States Sutton& : Hobson

. Not surprisingly, the principal typical output pattern is illustrated in figure 7.Sa mechanismfor making transitions from one loop to the other was to provide an external input closely resemblinga memory pattern in the second loop. As the amount of aminergic modulation was decreased , partially charac the model NREM state within a , but with , terizing sequencing loop persisted decreased stability . Relatively small perturbations disrupted the preset transitions . The instability of sequences at the level of the entire network was correlated with fluctuations at the synaptic and membrane levels. As the aminergic system turned off, there was increasednoise in synaptic transmission , with corresponding increasesin membrane potential fluctuations. On average , a larger proportion of neurons remained closer to their firing thresholdswhen there was aminergic withdrawal. With further aminergic decreasesand the disinhibition of phasic cholinergic effects on the network, a spectrum of interrelated phenomenawas observed . This included (a) sequences that bifurcated between loops; (b) a to apparentdisorder,1 (c) a change changefrom orderly intraloop sequencing from apparent disorder to orderly progression within a single loop (" defib" rillation effectd ) a changefrom a disorderly pattern to another disorderly . An pattern; and (e) no or little effect on identifiable intraloop sequencing example of transition types (b) and (c), with the overall effect of inducing a bifurcation between the two loops, is shown in figure 7.sb. In general , lower-intensity, longer-duration PGO bursts were more effective in inducing bifurcations than higher-intensity, shorter-duration bursts. The networks also appearedto be surprisingly sensitive to the timing of the PGO-type bursts in determining the fonn of the transitions. PGO-induced bifurcations were possible in all states , including the simulated waking state , if the bursting was intense , of long duration, and occurred at selective times with respect to the microscopic dynamics . This finding had an analogy with the shifts in attention correlatedwith PGO bursting during the startle responsein wakefulness . Finally, the mechanismof phasic excitatory facilitation, via PGO bursts , was not unique in inducing a plethora of network . For example , superimposing phasic aminergic withdrawal responses on a state of tonic aminergic decreaseled to results similar to those just described . In summary , the findings suggestedways whereby networks dramatically alter their sequencingproperties through changesin modulation. The results do not prove that combined aminergic and cholinergic changes mediate bifurcations in visual imagery, but they are certainly consistent with the notion that neurophysiological state changesalter global process es in largescale networks. Different processing states not only exist within the same network but relatively small changes in neuromodulation can induce significant state transitions over widespread regions. In biological networks, es. widespreadregions are linked to cognitive process

180

Behavioral States

. , ~ ! I \ . I . ! _ . . . . ~ ~ ~ , ! I . . 8 . r ; r ; . . . 8 a = . . ~ ~ ~ ~ , ~ s , ,
IIIIIIIIII ~

IS

. ~

"

n'

..

lS

Sutton & : Hobson: Waking and SleepingStates

' depaAo V wa a -

"

a aq o

Learning
It is generally acceptedthat memory and learning are interrelated process es , and that sleep is critical homeostatic mechanismin normal cognition. We have extended our theoretical approachto include state - dependentlearning modifiable in the by including synapses sequencing network. A central assumption in doing this was that basic synaptic mechanismsunderlying . However, their plasticity were essentially the same across different states local efficienciesmay be decreaseddue to the withdrawal of the biogenic amines , especiallyNE, in sleeprelative to waking. Our results are preliminary. We implementedstandardbiophysical mechanisms -dependent synaptic changes to incorporate featuresof use , including term enhancement and or Brown Kairiss , , , & ( long depression forgetting Keenan , 1990). The general mathematical formulation is included in the , we have found that the REM appendix. While more simulationsare required sleep state is particularly well suited for certain types of computation, such as rapid scanningand course learning. In the sequencingmodel, the system had the capacity to move readily between apparently uncorrelatedpatterns , whereas , in the waking and NREM sleep states , it did not spontaneouslydo this. When coupled with a relatively withdrawn, but nevertheless present , source of local plasticity, the possibility of novel memory formation, via unsupervisedlearning, was abundantly clear in the REM sleep state. The network appearedto be computing by superimposingphasicand tonic noise . In making transitions out of upon an otherwise highly constrained system the REM sleepstate back into the NREM sleepand waking states , the learning becamemore supervisedand efficaciousin responseto external inputs.

DISCUSSION
In this chapter , we have outlined an approach to investigating sleep and cognition using artificial neural networks . While the models are impres sionistic , they nevertheless allow computational studies of the intricate relationships between neuromodulation , state- dependent cognition , and plasticity . Our studies support the hypothesis that different neurophysio logical states mediate different types of sequencing and learning operations , and they suggest that an evolution of state changes may enhance the overall learning capability and efficiency of biological neural networks . Our work complements the research of other investigators exploring the seemingly hidden domain of the sleeping brain . We are optimistic that newer techniques , such as functional MRI , combined with neural network methods , electrophysiology , and subjective reports , will begin to weave an integrated story of neural computation in human waking and sleep.

182

BehavioralStates

ACKNOWLEDGMENTS
We acknowledge support from the McDonnell -Pew Program in Cognitive Neuroscience , the MacArthur Foundation and the National Institutes of Health, MH grants ROl13923 and K21 01080.

NOTE
1. The apparentdisorder revealedseveralsequences in loops A and B running out of phasewith relative delays . These delays were generally less than the time constant for intermemory transitions within a single loop.

REFERENCES
Abeles ., Tishby : Vaadia , M., Bergman , H., Gat , I., Meilijson , I., Seidemann , E , N., & . E. (1995 ). -stationarystates Corticalactivity Ripsamongquasi . Proceedings the National of Academy of Sciences - 8620 . , 92 , 8616 : VanderMaelen of the reticular . In core : Serotonin , G., & , C. (1986 ). Specific systems Aghajanian - The nervous - 256 V. Mount castle(Ed .), Handbook , Vol. 4 (pp. 237 ). of physiology system . Bethesda . MD: American Society Physiological Amit, D. J. (1989 brain . Cambridge . . : Cambridge , U.K ). Modeling function UniversityPress Anderson : Rosenfeld .). (1988 . Cambridge . , J. A., & , E. (Eds , MA: MIT Press ). Neurocomputing : Comparison Antrobus of word frequencies , J. S. (1983 ). REMandNREMsleep by cognitive reports - 568 classes . Psychophysiology . , 20 , 562 : Kleitman , E., & , N. (1953 ). Regularly Aserinsky occurring periodsof ocularmotility and concomitant - 375 . Science . , 118 , 361 duringsleep phenomena Baars : Banks .). (1994 andcognition . NewYork: Academic . Press , B. J., & , W. P. (Eds ). Consciousness Bear : Singer and , M. F., & , W. (1986 ). Modulationof visualcorticalplasticityby acetylcholine - 176 noradrenaline . Nature . , 320 , 172 Bootzin . R., Kihlstrom : Schacter andcognition . Washington ,R , J., F, & , D. L. (1990 , DC: ). Sleep American Association . Psychological Bowker : Morris on, A. R. (1976 . BrainResearch , R. M., & , 102 , ). ThestartlereflexandPGOspikes - 190 185 . .). (1987 book . Oxford . Brook , S. (Ed , U.K.: Oxford UniversityPress ). TheOxford of dreams . Archives C., & : Bizzi BrooksD . , E. (1963 ). Brainstemelectrical activity during deepsleep ltaliennes deBiologie - 665. , 101 , 648 : Biophysical : Keenan mechanisms Brown , T. H., Kairiss , E. W., & , C. L. (1990 ). Hebbian synapes - 511 . Annual . andalgorithms Review , 13 , 475 of Neuroscience : a role for noisybrainstates Buzsaki - stagemodelof memorytraceformation . , G. (1989 ). Two - 570 . Neuroscience , 31 , 551 . , Baghdoyan : Hobson - geniculo - occipital , C. W., Lydic ,R , H. A., & , J. A. (1987 ) . Ponto Callaway waves : Spontaneous visualsystemactivationoccurringin REM sleep . Cellular and Molecular - 149 . ,7 , 105 Neurobiology : Computer : Winston simulation of Clark , J. W., Rafelski , J., & , J. V. (1985 ). Brainwithout mind - 273 . Physical neural networks with modifiable neuronal interactions . , 123 (4), 215 Reports

: WakingandSleeping Sutton& I: Hobson States

Collins : Imhoff , J. J., Chow , C. C., & , T. T. (1995 resonance without tuning ). Stochastic . Nature , 376 - 238 . , 236 - 114 : Mitchison Crick . F., & , G. (1983 of dream . Nature ). Thefunction . , 304 , 111 sleep Datta . (1995 ,S ). Neuronal area : Relationship to behavioral statecontrol activityin the peribradlial . Neurosdence andBiobehavioral Reviews , 19 ( 1), 67- 84. Davis : Schoenfeld , M., Jacobs . I. (Eds , B. L., & .). (1989 ,R vertebrate ). Modulation neural of defined drcuits . NewYork: New York Academy of Sciences . Dement : Kleitman , W., & , N. (1957 in EEGduring sleepand their relation ). Cyclic variations to eye movements , body mobility and dreaming . Electroencephalography and ClinicalNeuro - 690. , 9, 673 physiology Fishbein : Chattman , W., Kastaniotis , C., & , D. (1974 : Prolonged ). Paradoxical sleep augmentation . Brain Research , 79 , 61- 77. followinglearning Flicker : Hobson , C., McCarley , R. W., & , J. A. (1981 neurons : State ). Aminergic controlandplasticity - 166 in threemodelsystems . Cellular andMolecular . ,1 (2), 123 Neurobiology Foulkes . : A , D. (1985 ) Dreaming cognitive . Hillsdale : Erlbaum . , NJ psychological analysis Freeman in thenervous , W. J. (1975 . NewYork: Academic ). Massaction Press . system Freud , S. (1895 ). Projectfor a scientific . In Standard edition psychology of thecomplete psychological - 397 worksof Sigmund Freud .), Vol. 1 (pp. 281 , trans (J. Strachey : Hogarth ). London , - 1974 1953 . . (1900 Freud ,S ). The .). NewYork: Avon 1965 (J., Strachey , trans . interpretation of dreams Geszti : pazmandi , T., & within bounds , F. (1987 anddream ). Learning . Journal A: of Physics sleep - L1303 Mathematical andGeneral . , 20 , L1299 : Borbely Gillin , J. C., & : A neurobiological , A. A. (1985 windowon affective ). Sleep disorders . - 542 Trends in Neuroscience . , 8, 537 Gutfreund : Mezard , H., & of temporalsequences , M. (1988 in neuralnetworks ). Processing . - 238 Review Letters . , 61 , 235 Physical Harris -Warrick : Marder . (1991 , R. M., & of neural ,E networks ). Modulation for behavior . Annual Review , 14 , 39- 57. of Neuroscience Hasselmo , M. E. (1993 and learningin a corticalassociative ). Acetylcholine . Neural memory , 5, 32- 44. Computation " " Hinton -sleep , G. E., Dayan : Neal , P., Frey , B. J., & , R. M. (1995 ). The wake algorithmfor neural networks . Science 268 1158 . 1161 , , unsupervised brain . NewYork: Basic Hobson , J. A (1988a ). The Books . dreaming Hobson , J. A. (1988b and heteroplasticity : Functional ). Homeostasis of behavioral significance statesequences - 220 . In R . Lydic & : J. F. Biebuyck .), Clinical (Eds (pp. 199 physiology of sleep ). Bethesda , MD: American . Society Physiological Hobson , J. A. (1990 ). Activation , input source . In R. R . Bootzin , andmodulation , J. F. Kihlstrom , & : D. L. Schacter .), Sleep andcognition (Eds (pp. 25- 40). Washington , DC: American Psychological Association . Hobson , J. A , Hoffman : Kosner , S. A., Helfand , R., & , D. (1987 and the ). Dreambizarreness -synthesis - 164 activation . Human . , 6, 157 hypothesis Neurobiology Hobson , J. A., Mamelak : Sutton , A. N., & , J. P. (1992 : Must fit dimensions ). Modelswanted of anddreaming . In JE . Moody : R. Uppmann , S. J. Hanson .), Advances ,& in neural sleep (Eds information 4 (pp. 3- 10 Mateo . ). San , CA: MorganKauhnann processing systems

184

BehavioralStates

-stategenerator : An activation : McCarley Hobson , J. A., & , R. W. (1977 ). The brainasa dream . American oumRl 134 1335 1368 . of the dream , , ] of Psychiatry synthesis process hypothesis W. W. : reciprocal A. R . & : P. . oscillation discharge . 1975 , , Wyzinki , Hobson , J , McCarley ( ) Sleep cycle brainstem neuronal . Science 189 55 58. two , , by groups . In V. B. Mountbasis of behavioral state control : Steriade Hobson , J. A , & , M. (1986 ). Neuronal - Thenervous - 823 .), Handbook castle , MD: , Vol 4 (pp. 701 ). Bethesda (Ed systmz of physiology . American Society Physiological . G. (1983 hasa stabilizing effectin collective : Palmer , J. J., Feinstein , D. I., & ,R ). Unlearning Hopfield . Nature - 159 . memories , 304 , 158 -termpotentiation enhancement of long : Johnston . (1988 , W. F., & , D. J ). Noradrenergic Hopkins - 687 2 . in the hippo . oumRl 667 at mossy fibersynapses , 59 ( ), ] of Neurophysiology campus and : Archer , T. (1981 , T. U. C., Callenholm , N. E. B., Mahammed , A D., & ). Noradrenaline Jarbe - 501 -dependent . . Physiology andBehavior extinction effect the context , 495 , 38 D. L., Kandel ., Klein , S., Hawkins , R., Glanzman , M., Hochner , B., Shuster , M., Siegelbaum , E. R : New insights into andlearning : Abrams modulation Castellucci , V. F., & , T. W. (1987 ). Synaptic W. W. M. E. Gall & : from the of behavior . In G. M. Edelman transmission , , study synaptic .), Synaptic Function . NewYork : Wiley. Cowan (Eds ., & : Ungerleider Karni ,R , L G. (1995 , P., Adams , M. M., Turner ). Functional , A , Meyer , G., Jezzard . Nature MRI evidence for adult motor cortexplasticityduringmotor skililearing , 377 , - 158 155 . on : Sagi KamiA , Tanne , J. J. M., & , D. (1994 , D., Rubenstein , B. S., Askenasy ). Dependence a skill . Science 265 679 682. of REMsleep of overnight , , improvement perceptual : Thresholds ., & Keeler : Ross , hysteresis , , J. (1989 , J. D., Pichler , E. E ). Noisein neuralnetworks -to-noise . Proceedings of signal and neuromodulation , 86 , of Sciences of the NationalAcademy - 1716 . 1712 . Proceedings stategeneration Kleinfeld , D. (1986 ). Sequential of the by modelneuralnetworks 83 9469 9473 . National Sciences , , Acadmzy of -regulatory . In W. P. Koella & : P. Levin function of sleep Kramer : Roth , M., & , T. (1972 ). Themood - 570 : Karger . .), Sleep (pp. 563 ). Basel (Eds -Ede : Weitzman . E., Czeisler Kronauer , M. C., & , E. D. (1982 , S. F., Moore ). ,R , C. A , Pilato . Amm " can with two interacting oscillators circadian Mathematical modelof the human system . , 242 (11 ), RJ- R17 ] oUmRl of Physiology . Mathematical in the brain Biosciences of persistent states Little , 19 , , W. A. (1974 ). Theexistence - 120 101 . -like statemodulated . In asan oneiric Uinas . R., & : Ribary , U. (1994 ,R ). Perception by the senses -scale . neuronal theories .), Large C. Koch& : J. L Davis(Eds , MA: MIT Press of thebrainCambridge . New York : DoverPress . Lotka , A (1956 ). Elements biology of physical asthe cognitivecorrelate of altered bizar Tenes : Hobson Mamelak , A. N., & , J. A (1989 ). Dream - 222 . Neuroscience . ] oumRl neuronal behavior in REMsleep ,1 (3), 201 of Cognitive . , von Frenke U : Frank ,R , R., & , G. , B., Franco , G., Poirrier , P., Dive, D., Salmon , E., Sadzot Maquet utilizationduring sleep (1990 ). Cerebral by positron glucose - wakecycle in man detennined -D . method . Brain Research - 143 - 2- deoxy - glucose and[18F emission , 513 , 136 ]2-Auoro tomography overthe sleep : A structural : Hobson , J. A (1975 , R. W., & ). Neuronal excitability cycle McCarley . Science model andmathematical , 189 , 58- 60.

andSleepingStates : Hobson : Waking Sutton&

: Hoffman , E. (1981). REM sleepdreamsand the activation-synthesishypothesis McCarley, R. W., & . American , 138 , 904- 912. Journalof Psychiatry . W., & : Massaquoi , S. G. (1986). A limit cycle mathematicalmodel of the REM sleep McCarley, R osdllator system . American . , 251, R1011- R1029 Journalof Physiology : Duffy , F. H. (1983). Human cerebralpotential associated , J. W., & McCarley, R. W., Winkelman with REM sleep rapid eye movements : Links to PGO waves and waking potentials. Brain Research , 274 , 359- 364.

McClelland : O' Reilly , J. L., McNaught on , B. L., & arecomplementary , R. C. (1994 ). Why there in the hippo and neocortex : Insights andfail. uresof con learning systmrs campus from the success nectionist models andmmlory . CMU Technical No. PDP .CNS .94.1. of learning Report . Cambridge McCullochW. S. (1965 . ). Embodiments , MA: MIT Press of mind ., & : Pitts McCullochW. S of the ideas , W. (1943 immanent in nervous ). A logicalcalculus activity - 133 . Bulletin . , 9, 115 of Mathematical Biophysics -Schott Merritt, J . M., Stickgold : Hobson , R., Pace , E., Williams , J., & , J. A (1994 ). Emotion profiles in thedreams of menandwomen . Consciousness andCognition , 3(1), 46- 60. : Godbout .). (1990 andbiological , J., & : Basic , R. (Eds mechanisms and ). Sleep Montplaisir rhythms to psychiatry . Oxford . , U.K : Oxford UniversityPress applications MooreR. Y. (1979 of central neural mechanisms ). The anatomy endocrine . regulating rhythms In D. T. Krieger .), Endocrine : Raven . (Ed Press (pp. 63- 87 ) . New York rhythms -Ede -wake Moore : Czeisler .). (1984 , M. C., & models , C. A. (Eds ). Mathematical of thecircadian sleep . New York : Raven Press . cycle Moss : Wiesenfeld of background , F., & , K. (1995 noise . Scientific American ). Thebenefits , 273 (2), 66- 69. Nakao : Yamamoto , M., Takahashi , T., Mizutani , Y., & , M. (1990 ). Simulation studyon dynamics transition in neuronal and symmetry neural activity duringsleep cycleby usingasynchronous - 250 networkmodel . Biological . , 63 , 243 Cybernetics Olivier : Poirer , A., Parent of the thalamic ,A,& , L. J. (1970 nucleion the basis of ). Identification theircholinesterase content in themonkey . Journal , 106 , 37- 50. of Anatomy Poe . , RectorD. M., & : Harper . M. (1994 , G. R reflected ,R ). Hippocampal opticalpatterns during - 2942 andwakingstates in thefreelybehaving cat . Journal . , 14 (5), 2933 sleep of Neuroscience : Frost and localcerebral Ramm , P., & , B. J. (1986 metabolism in the cat during slow ). Cerebral - 124 waveandREMsleep . BrainResearch . , 365 , 112 Recanzone , G. H., Merzenich : Dinse . (1992 , M. M., Jenkins , W. M., Grajski , K A., & , H. R ). Journal - 1056 . , 67 , 1031 of Neurophysiology Rechtschaffen , A , Bergmann : Gilliland , C., Kushida , C. A , & , M. A. (1989 ). Sleep , B. M., Everson in the rat: X. and discussion of the . 12 68 87. , , deprivation Integration findingsSleep - 787 Reiss : Taylor . G. (1991 , M., & . Neural ,J Nehoorks . ). Storingtemporal , 4, 773 sequences : Dement , J. N., & of the human , W. C. (1966 ). Ontogenetic Roffwarg , H. P., Muzio development . Science - 619. , 152 , 604 cycle sleep - dream -timeattentional : Moore . N. A , & models for classical , J. W. (1985 Schmajuk ). Real conditioning - 290 in thehippo . . , 13 , 278 Physiological Psychology campus of actionof noradrenaline in thebrain , M. (1985 . In B. WiD ). Mechanisms :R . , P. Schmitt Segal ,& - 239 Alford .), Brainplasticity andmemory (Eds , learning (pp. 235 ). New York: Plenum Dalrymple . Press

186

BehavioralStates

Selverston and behavior . New York: PlenumPress . , A. I. (Ed). (1985). Model neuralnetworks -Schreiber Servan : Cohen : , Printz, H., & , J. D. (1990). A network model of catemolamine effects -to-noise ratio, and behavior. Science Gain , signal , 249, 892- 895.

. B. (1995 Silberstein of neocortical . In P. t . Nunez(Ed .), Neocor ,R ). Neuromodulation dynamics Hcal andhuman EEG - 627 .: Oxford UniversityPress . (pp. 591 , U. K ). Oxford dynamics rhythms . , Valatx Smith : Jouvet increases in paradoxical , C., Kitahama , K , J. t ., & , M. (1980 ). Prolonged and after avoidance task . 3 1 , ( ), during Sleep 67- 81. sleep acquisition -nerveactivity Somers . , Dyken : Abboud , V. K , M. E., Mark , A L, & , F. M. (1993 ). Sympathetic - 307 in normal . NewEngland . , 328 , 303 subjects ]oumRl of Medicine duringsleep : Kanter in asymmetric neuralnetworks . , H., & , I. (1986 ). Temporalassociation Sompolinsky - 2864 Reviews Leftm . , 57 , 2861 Physical Steriade : Uinas . R. (1988 states of thethalamus andassociated neuronal , M., & ,R ). Thefunctional . Physiological Reviews - 742 . , 68 (3), 649 interplay Steriade : McCarley control andsleep . New York : , M., & , R. W. (1990 ). Brainstem of wake fulness Plenum Press . Steriade : Sejnowski osci Uations in the , M., McCormick , D. A , & , T. J. (1993 ). Thalamocortical andaroused brain . Science - 685. , 262 , 679 sleeping Sutton : Hobson - dependent andlearning . In F. H. Eeckman , J. P., & , J. A (1994 ). State sequencing - 280 in neurons andneural : Kluwer Academic Publisher . (Ed), Computation (pp. 275 ). Boston systems Steriade M. Pare D. Parent A Y. & : . of cholinergic and non , , , , , , Smith , (1988 ) Projections neurons of the brainstem coreto relayand associational thalamic nucleiin the cat cholinergic andmacaque . Neuroscience , 25 , 47- 67. monkey -wakecycle structure . New York : , S. H. (1986 ). Themathematical of the human Strogatz sleep -Verlag . Springer - 208 Sutton andcomputing . Information Sciences . , J. P. (1995 ). Neuroscience , 84 , 199 algorithms Sutton : Hobson of wakingandsleeping . , J. P., Mamelak , AN ., & , J. A. (1992a ). Modelingstates - 143 Annals . , 22 (3), 137 Psychiatric Sutton . A. (1992b : Hobson - dependent , J. P., Mamelak , AN ., & , J ). Network modelof state . In . . R . P. S. & : . Advances in neural , Hanson , JE J (Eds ), Moody information sequencing Lippmann - 290 4 (pp. 283 Mateo : MorganKaufmann . , CA ). San systems processing -Schott Sutton ., & : Hobson , J. P., Rittenhouse , C. D., Pace , E., Merritt, J. M., Stickgold , R , J. A : A narrative . Consciousness (1994a ). Emotionand visualimageryin dreamreports graphingapproach andCognition , 3(1), 89- 99. -Schott Sutton ., & : Hobson , J. P., Rittenhouse , C. D., Pace , E., Stickgold , R , J. A. (1994b ). A new to dream bizarreness : and of visual in attention continuity discontinuity approach Graphing narrative . Consciousness , 3(1), 61- 88. reports .andCognition Takahashi andgenetics of circadian . , J. S. (1995 ). Molecular rhythmsin mammals neurobiology - 553 Annual Review . , 18 , 531 of Neuroscience -ViUars surla theorie dela luftepourla vie . Paris : Gauthier . Volterra , V. (1931 ). Lefons mathtmatique WiUiams : Hobson in dreams andfantasies , J., Merritt, J., Rittenhouse , C., & , J. A. (1992 ). Bizarreness - 185 -synthesis : Theactivation . andCognition Consciousness . ,1 , 172 hypothesis Wilson : McNaught of the hippocampal ensemble codefor , M. A., & on , B. t . (1993 ). Dynamics - 1058 . Science . , 261 , 1055 space

187

States : Hobson : WakingandSleeping Sutton&

: McNaught Wilson , M. A , & of hippocampal on , B. L. (1994 ensemble memories ). Reactivation . Scima - 679 . , 265 , 676 duringsleep Wimbauer : vanHemmen , 5., Klemmer , N., & of unlearning . Neural , J. L. (1994 Networks ). Universality - 270 . ,7 (2), 261 Winson of dreams . Scientific , J. (1990 American ). Themeaning , 263 (5), 86- 96. Wu , J. C., Gillin , J. C., Buchsbaum T, HazlettE., 5icotle , M. S., Henhey , WU , Gillin , Bu , Hesshey , N., & : Bunney , W. E. Jr. T., (1991 on cerebral ). Theeffectof sleep metabolic deprivation glucose - 162 ratein normal humans assessed with positronemission . Sleep . , 14 , 155 tomography

APPENDIX
Mathematically, the temporal sequencingnetwork consistsof N model neurons . Eachneuron is assigneda value Si(t ) = :t: I , i = 1, . . . , N , corresponding to whether or not it is firing at time t. The neurons are linked together (1 ) and (2), , Iii by two kinds 0f synapses Iii i ~ j,

l)= f.ff. Ji } ), t ~ = 1 p

(1)

-l + q 2 lfa ) = J ; faf } ~L = 1 ), p

encodesa set of p uncorrelatedpatterns { f.f }~ l ' J .l = 1, . . . , p, where eachf.f -stable has the value :I: 1 with equal probability . These patterns are the quasi . . . = + + + memories . Sequentialtransitions among memoriesJ 1 2 .l q< p are inducedby

(2)

where 1 is a relative weight term . The average time spent in a memory pattern is 'C . The membrane potential is given by

N l 2 )Sj )Sj = t hi ( ) E Ui (t) +Ji (t - -r)] +bi (t) + 1i (t). ~ ~ = jl

(3)

The two terms contained in the brackets reflect intrinsic network interactions , while phasic PGO effects are given by 15 , non-fCC , ;(t ) . External ;(t ). Dynamic evolution follows the synchronous inputs are denoted by 1 rule updating (4) the amount of aminer lJj(t) is the membranethreshold and p parameterizes gic like modulation. Synaptic learning in equations(1) and (2) has the form k = 1, 2. (5) F is a function of presynaptic and postsynaptic firing , and e gives the learning rate, which is maximal in the simulatedwaking state and minimal during the model REM sleepstate . Sj(t + 1) = :t 1, with probability

= l=.8 ;(I)- O ;(I)I} - I , [It {I + e 2

k k ) ) = Ji ) (t),Sj (t)], Ji [Si ~ (t + 1 ~ (t) + eF

188

Behavioral States

III

Neuropsychological Tests and Clinical Syndromes

Stroop Task , Language , and : Models of Cognitive Neuromodulation Deficits in Schizophrenia


David Servan - Schreiber and JonathanD . Cohen

Ultimately, a complete understanding of human behavior must derive from an understandingof the biological mechanisms upon which it relies. Despite this obvious, and intimate link between behavior and biology , researchin each of these domains has remained relatively autonomous particularly when it concernspsychopathology. For example , schizophreniais markedby a wide variety of behavioral deficits , including disturbancesof attention, language and , problem solving. At the sametime, schizophreniais processing characterizedby biological abnormalities , including disturbancesin specific neurotransmitter systems (e.g., dopamine and norepinephrine ) and ana tomical structures(e.g ., the prefrontal cortex, or PFC , and the hippocampus ). However, the behavior and biology of schizophreniahave remainedseparate fields of inquiry . Despite a modem consensusthat information-processing deficits in schizophreniaare the result of underlying biological abnormalities , few efforts have been made to specify exactly how these phenomenarelate to one another. In this chapter we draw upon the framework of connectionist models to addressthis gap between behavior and biology . The modeling framework provides theoretical concepts that are intermediate between the details of neuroscientificobservations and the box-and-arrow diagrams of traditional . In particular, we information processing or neuropsychological theories models to the of connectionist ability explain aspects of schizophrenic explore . At behavior in terms of specific underlying biological disturbances the behavioral level, the models addressboth normal and schizophrenicperformance in three experimental tasks : two that tap attentional performance (the Stroop task and the continuous performancetest), and one that measures -processingabilities (a lexical disambiguationtask). The models use language a common set of information-processingmechanisms , and show how a number of seeminglydisparateobservationsabout schizophrenicbehavior can all be related to a single functional deficit: a disturbance in the processing of context. The models also suggest that this functional deficit may be explained : a reduction in the effects of the neurotransmitter by a specificbiological disturbance . First, we show how a particular parameter dopamine in the PFC

of the models can be used to simulate the neuromodulatory effects of dopamine at the neuronal level. We then present the results of simulations in which this parameter is disturbed within a module corresponding to the function of the PFC . In eachof the three behavioral simulations , this disturbance leads to changesin performance that quantitatively match those observed for patients with schizophreniain the corresponding tasks . These findings suggestthat a number of the disturbancesof attention and language found in schizophreniamay all result from a disturbancein the processingof context which, in turn, may be explained by a single biological abnormality, a reduction of dopaminergicactivity in the PFC . The backgroundof the models presentedin this chapter spansa large and diverse literature concerningcognitive deficits in schizophrenia , the anatomy and physiology of dopamine systems , the neurophysiology and neuropsychology of frontal cortex, and the role of these biological systems in . A full review of these data is beyond the scope of this communication schizophrenia -Schreiber (seeJ. D. Cohen & Servan , 1992, for a more comprehensive review). Here, we highlight five empirical observations concerning information processingand biological deficits in schizophrenia . The purpose - which of our simulations is to show how these observations range from the biological to the cognitive level- can be articulated within the same model and account for specificaspectsof behavior. These observations are: (1) The performanceof patients in a variety of cognitive tasks indicates a decreasedability to use context information for selecting appropriate behavior . By context , we meaninformation that is relevant to, but is not part of, the content of a behavioral response . This can be task instructions or specific previous stimuli that determine correct behavior. (2) The PFC is directly involved in, and necessaryfor, the representationand maintenanceof context information. (3) The normal function of the PFC relies on the activity of the mesocorticaldopaminesystem . (4) Dopamine has a modulatory effect on the activity of the PFC by influencing the responsivity, or gain, of cells in this brain region. (5) Schizophreniais associatedwith abnormalitiesof both frontal cortex and dopamineactivity .

DISTURBANCES IN THEPROCFSSING OF CONTEXT The Stroop Task


This task taps a fundamentalattentional phenomenon : the ability to respond to one set of stimuli, even when other, more compelling stimuli are available . The paradigm consists of two subtasks . In one, subjectsname the color of the ink in which a word is printed. In the other, subjectsread the word aloud while ignoring ink color. Normal subjectsare less able to attend selectively to colors (ie ., ignore words) than the reverse . If patients with schizophrenia suffer from an attentional deficit, then this effect should be exacerbated ; that is, they should be less able to ignore word information, and should show a

192

Tests andClinical Neuropsychological Syndromes

greater interferenceeffect. This prediction is supported by studies of performance by patients with schizophreniain the Stroop task (Abramczyk, Jordan , & : Hegel, 1983 : Krus, 1960). However, because an overall slowing ; Wapner & of reaction time is also observed , the significanceof an increaseininterference has been called into question : This may simply reflect an unanticipated effect of general slowing of performance , rather than of a specificattentional deficit. This issue has not been resolved in the literature. Below, we show how a simulation model of this task can help distinguish the effectsof ageneral slowing from those of a specificattentional deficit. Considerationsof the Stroop effect typically focus on the role of selective attention. However, the processingof context is also central to this task. In order to respond to the appropriate dimension of the stimulus , the subject must hold in mind the task instructions for that trial. These provide the necessarycontext for interpreting the stimulus and generating the correct . In Stroop experiments , trials are typically blocked by task (e.g ., all response color naming, or all word reading), so that the proper context is consistent , - such as the and regularly reinforced . However, in other attentional tasks continuousperformancetest (Cpn - this is not the case .
The Continuous Performance Test

In the CPT, subjectsare askedto detect a target event among a sequence of stimuli and to avoid to distractor stimuli. The , briefly presented responding target event may be the appearanceof a single stimulus (e.g., detect the letter X appearing in a stream of other letters), or a stimulus appearing in a particular context (e.g ., respond to X precededby A ). Patients with schizophrenia (and often their biological relatives ) are typically impaired in their to discriminate between target and distractors on this task , compared ability with normal and patient controls (e.g ., Kometsky, 1972; Nuechterlein , 1984). This deficit is most apparentin versions of the task that make high processing " " demands . For example , in the CPT-double a target event consists of two consecutiveidentical letters. Memory for the previous letter provides the context necessary to evaluate the significance of the current letter; . Patientswith schizophrenia inability to use this context would impair performance in this and similar versions of the task. perform especiallypoorly

Schizophrenic LanguageDeficits
Patientswith schizophreniaalso show poor use of context in languageprocessing . Chapman , Chapman , and Miller (1964) first demonstratedthis in a ' . They found that study of thesepatients interpretation of lexical ambiguities the patients tended to interpret the strong (dominant) meaning of a homonym used in a sentence , even when context suggested the weaker (subordinate " . For ) meaning , given the sentence The farmer needed a example " new pen for his cattle, the patients interpreted the word " pen" to mean

193

-Schreiber Servan & t Cohen : C~ nitiVE Deficits in Schizophrenia

. They did not writing implement more frequently than did control subjects differ from control subjects in the number of unrelated meaning responses that were made (e.g ., interpreting " pen" to mean " fire truck" ), nor did they differ in the number of errors made when the strong meaning was correct. , we tested the idea that patients with schizophreniaare restricted in Recently the temporalrangeover which they can process linguistic context (Cohen, -Schreiber Servan , Targetal ., 1992). We designed a lexical ambiguity task , similar to the one used by Chapmanand his colleagues , in which we could involved. manipulatethe temporal parameters Subjects were presented with sentencesmade up of two clauses ; each clauseappearedone at a time on a computer screen . One clausecontainedan " " ambiguous word in a neutral context (e.g ., you need a PEN ), while the other clause provided disambiguating context (e.g ., " in order to keep " " ". chickens or in order to sign a check ) Clauseswere designedso that they could be presentedin either order: context first or context last. The ambiguity in each sentencealways appearedin capital letters, so that it could be identified by the subject . Subjectswere presented with the sentencesand, each were asked to interpret the meaning of the ambiguity as it , following was usedin the sentence . Sentences were distributed acrossthree conditions: a weak correct , context last () ; (b) weak meaningcorrect, context first; meaning (c) strongmeaning correct, context first. For example , a given subject would have seenthe ambiguity " pen" in one of the three following conditions, and then have chosenhis or her responsefrom the list of possiblemeanings : ' , context (a) you cant keep chickens(weakmeaning first) [clearscreen ] / pause without a PEN or , conterllast) (b) without a PEN (weakmeaning clear screen [ ] /pause ' you cant keep chickens or (c) ' , context you cant sign a check(strongmeaning first) [clearscreen ] /pause without a PEN [clearscreen ] / pause

The meaning of the word in capital letters is: a writing implement (strongmeaning ) a fencedenclosure(weakmeaning ) a kind of truck (unrelated ) meaning The results of this study (shown in figure 8.7A ) corroborated both the original findings of Chapmanet al., and the explanation of their findings in

194

a J Tests andClinicalSyndrome ! Neuropsychologic

terms of an inability to maintain context. Thus, as the authors found, patients with schizophreniamade significantly more dominant meaning errors than did controls when the weak meaning was correct. However, this only occurredwhen the context camefirst (i.e., lessrecently- condition [b] above) and therefore had to be maintained until the time when the ambiguous " " . When context came last (i.e., more recently), patients ( pen ) is processed . This was the only type with schizophrenia correctly chosethe weak meaning of error that reliably distinguished thesepatients from controls. These findings suggest that the impairment observed in language tasks :a may be similar in nature to the impairments observed in attentional tasks in and context over time. difficulty representing maintaining PREFRONTAL CORTEX, CONTEXT, AND DOPAMINE Several studies suggest that frontal areasare specifically involved in maintaining context information for the control of action. For example , at the neurophysiological level, Fuster (1997) and Goldman Rakic (1987) have observed cells in the PFC that are specific to a particular stimulus and . They have , and that remain active during a delay between these response in the of are maintained PFC which that neural activity argued patterns to a . At the information needed encode the temporary guide response behavioral level, these authors and Diamond (e.g ., 1989) have also reported data showing that the PFC is needed to perform tasks involving delayed responsesto ambiguous stimuli. Diamond has emphasizedthat prefrontal , in particular, to overcome reflexive or previously reinforced memory is required responsetendenciesto mediate a contextually relevant but otherwise . weaker response Furthermore , it has been shown that dopaminergicinnervation of the PFC is necessary for this brain region to maintain contextual information. Experimental lesions in animals , to this dopaminergic , or cIinicallesions in humans to the itself on behavioral tasks mimic the effect of lesions PFC can supply Rosvold . Brozoski Brown for context e. , , , et al., 1979). , ( g requiring memory

NE UROMOrn JLA .TORY EFFECTS OFDOPAMINE


Several anatomical and physiological observations support the idea that catecholamines such as dopamine and norepinephrine modulate information processing in the brain . Dopamine and norepinephrine neurons originate in discrete nuclei localized in the brain stem and their fibers project radially to several functionally different areas of the central nervous system (CNS ). The baseline firing rate of these neurons is low and stable, and the conduction velocity along their fibers is slow . These characteristics result in a steady state of transmitter release and relatively long -lasting postsynaptic effects that are consistent with a modulatory role . Most important , recent evidence suggests that the effect of dopamine release is not to directly increase or

-Schreiber& : Cohen: Cognitive Deficits in Schizophrenia Servan

reduce the firing frequency of target cells (e.g ., Chiodo & Berger , 1986). Rather , like norepinephrine , dopamine seems to modulate the response properties of postsynaptic cells such that both inhibitory and excitatory . This effect has been described responsesto other afferent inputs are potentiated ' as an increasein the " signal-to -noise ratio" of the cells behavior or " " an enabling of its response(e.g ., Foote, Freedman , & Oliver 1975).

PREFRONTAL CORTEX AND DOPAMINE IN SCHIZOPHRENIA


The behavioral data reviewed earlier concerning schizophrenic performance deficits indicate an insensitivity to context , and a dominant response tendency . This is consistent with evidence that schizophrenia is associated with frontal lobe impairment . Patients with schizophrenia show typical frontal lobe deficits on standard neuropsychological tests, including the Wisconsin Card Sorting test ( WCST ) (e.g ., Malmo , 1974 ) and the Stroop task (as described above ). In addition , imaging and electrophysiological studies suggest an atrophy and abnormal metabolism in the frontal lobes of schizophrenic : Franzen, 1974 ). Recent studies have even patients (e.g ., Ingvar & demonstrated abnormal metabolism in the PFC of schizophrenic patients , specifically during performance on tasks requiring memory for context such as the WCST and a variant of the CPT ( Weinberger, Berman, & : Zec , 1986 ; R.M . Cohen , Semple, Gross , et al., 1987 ). This work confirms that anatomical and physiological deficits of frontal cortex may indeed be associated with some behavioral deficits observed in schizophrenic patients . Frontal lobe dysfunction in schizophrenia fits well with the prevailing neurochemical and psychopharmacological data concerning this illness. The PFC is a primary projection area for the mesocortical dopamine system , a disturbance of which has consistently been implicated in schizophrenia (e.g ., Meltzer & : Stahl, 1976 ). In view of these findings , several authors have proposed that reduced dopaminergic tone in PFC may be associated with hypofrontality in schizophrenia , and may be responsible for several of the cognitive deficits that have been observed (e.g ., Weinberger & : Berman, 1988 ).

SUMMARY
We referred to evidence that patients with schizophrenia inadequately maintain context for the control of action ; that the PFC plays a role inmaintaining context ; that an intact mesocortical dopamine system is necessary for normal PFC function ; and finally , that the mesocortical dopamine system is affected in schizophrenia . Despite a growing recognition that these observations are related , no theory has yet been proposed which explains - in terms of causal mechanisms- the relationship between disturbances in the PFC and dopamine on the one hand , and behavioral de6cits on the other . In the remainder of this chapter , we present a set of connectionist models that sim-

196

Tests andClinical Neuropsychological Syndromes

ulate the perfonnanceof patients with schizophreniain the tasks described above. SIMULA nON OF THE PHYSIOLOGICAL EFFECTS OF DOPAMINE In the models , the action of dopamineis simulatedas a changein a parameter of the function relating a unit' s input to its activation value. To do so, we first assume that the relationship between the input to a neuron and the neuron's frequencyof firing can be simulatedas a nonlinear function relating the net input of a model unit to its activation value. Physiological experiments suggestthat in biological systemsthe shapeof this function is sigmoid, with its steepestslope around the baselinefiring rate (e.g ., Freeman , 1979). The same experiments also indicate that small increments in excitatory drive result in greater changesin firing frequency than equivalent increments in inhibitory input. These properties can be captured by the logistic function with a constant negative bias: Activation = ~ .net+bIU ' ) 1 + e- (gam

(Seefigure 8.1, gain = 1.0.) The potentiating effects of dopamine can be simulated by increasing the gain parameterof the logistic function. As figure 8.1 (gain = 2.0) illustrates , with a higher gain the unit is more sensitive to afferent signals , while its baseline firing rate (net input = 0) remains the same . Elsewhere , we have
1 .0 Gain = 2.0 ' " ' " ~ ~ . > ~ ~ 0.1 Gain = 1.0

0.8

0..

0.2

0.0

..

-4

.2

0 Net Input

of the gainparameter on the logisticactivation functionof an individual Figure 8.1 The influence unit. Note that in gain , with anincrease of thenet inputon theunit's activation , theeHect is increased is true with a decrease , while the reverse in gain . These eHects simulate the consequences of dopamine release on targetneurons in the CNS .

197

-Schreiber& Servan ~in Schizophrenia : Cohen: Cognitive Deficit

shown that sucha changein gain can simulatea number of different catecho laminergic effects at both the biological and behavioral levels (e.g ., Servan Schreiber : Cohen , Printz, & , 1990). In order to simulate the effect of a neuromodulator , we change gain for all units in the model that are assumed to be influencedby that equally neuromodulator . For example , the mesocorticaldopamine system has extensive . To model the action of dopamine in this brain projections to the PFC area , we change the gain of all units in the module corresponding to this area . In the models described below, decreaseddopamine activity in prefrontal cortex was simulated by reducing the gain of units in the module used to represent and maintain context. In all three models , simulation of of with was conducted by reducing gain performance patients schizophrenia from a normal value of 1.0 to a lower value in the range 0.6 to 0.7.

SIMULA nON OFTHE STROOPEFFECT


Elsewhere , we have described a connectionist model of selective attention that simulates human performance in the Stroop task (Cohen, Dunbar, & : McClelland, 1990). In brief, this model consistsof two processingpathways, one for color naming and one for word reading , and a task demand module that can selectively facilitate processingin either pathway (figure 8.2) Simulations are conductedby activating input units correspondingto stimuli used in an actual experiment (e.g ., the input unit in the color-naming pathway representingthe color red) and the appropriate task demandunit. Activation is then allowed to spread through the network. After training with back this leads to activation of the unit to , propagation output corresponding " " the appropriate response (e.g ., red ). During backpropagation learning , the network is trained on " reading words" (i.e., associating inputs in the word -processing pathway with the corresponding output unit while the " " word -reading task demand unit is active) ten times more frequently than it is trained on " naming colors." The asymmetry in the training regimen produces larger weights in the word -reading pathway than in the colornaming pathway, which underliesthe basicStroop effect. This simple model is able to simulate an impressive number of empirical phenomena associatedwith the Stroop task. It captures the three basic effects : (1) the asymmetry in speedof processingbetween word reading and color naming, (2) the immunity of word reading to the effects of color, and (3) the susceptibility of color naming to interference and facilitation from words- including the fact that the interference effect is larger than facilitation . The model also capturesthe influenceof practice on interferenceand facilitation effects , the relative nature of these effects (i.e., their susceptibility to training), responseset effects , and stimulus onset asynchrony effects (see et al. . Cohen , 1990) This model also exhibits behaviors that make it relevant to understanding schizophrenic disturbancesof attention, and their relationship to the pro-

198

NeuropsychologicalTests and Clinical Syndromes

RESPONSE
" green " " red "

Figure 8.2

are the output (response ) units.

cessing of context. The model shows how attention can be viewed as the effect that context has on selecting the appropriate pathway for responding. Here, context is provided by the task instructions. Thus, when subjectsare presentedwith conflicting input in two dimensions(e.g ., the word GREENin red ink), they respond to one dimension and not the other, depending upon the contextin which it appears(i.e., the task: color naming or word reading). If the &ontal cortex is responsiblefor maintaining this context, and if schizophrenia involves a disturbanceof &ontal lobe function, then we should be able to simulatethe performanceof schizophrenicpatients in the Stroop task , if by disturbing processing in the task demand module. More specifically &ontallobe dysfunction in schizophreniais due to a reduction in the activity of its dopaminergicinput, then we should be able to simulate this by reducing the gain of units in the task demandmodule. Figure 8.38 shows the results of such a simulation, in which the gain of only the task units was reduced . ; all other units were unperturbed This change in the context (task demand ) module produced effects similar

199

-Schreiber6 Servan : Cognitive : Cohen DeAdts in Schizophrenia

Color

INKCOLOR

RED GREEN Word WORD Reading TASK DEMAND Network architecture . Unit!;at the the bottom are units units atthe ,and input top
green
Naming

A c O . 7 """'" , : f / 0 3 Word Color Colo Word Colo Co Conf Readi Nami Read Nam
B ~ Word Color Color Conflict Reading Naming
Gain 0 Reduced (0.64 ) (1.0) . Nonnal Empirical data a Schlzoprhenlcs . Nonnal contois Cascade rate . 6 Reduced (0.0075 ) ... Nonnal (0.01 ) for normaland schizophrenic , and results Figure 8.3 (A ) Strooptaskperformance subjects &om simulations the unitsonly) and(0 the cascade (8) (taskdemand manipulating gainparameter rate(all units . Empirical dataarethe response times(in seconds ) in the network ) for stimuli in eachconditionof the Strooptask over threeempirical studies & : Krus , averaged , ( Wapner 1960 et aI ., 1983 ; Abramczyk : Sweet data are the numberof ; Wysocki& , 1985 ); Simulation for processing stimuliof eachtype comparison , scaled cycles (0.2) to facilitate required by a constant between these andtheempirical data .

to those observed in patients with schizophrenia : an increase in overall responsetime, with a disproportionate increasefor color-naming interference trials. Thus, the model shows that a lesion restricted to the mechanismfor processingcontext can produce both an overall degradation in performance as well as the expectedattentional deficit. The model also allows us to compare the effects of this specific disturbance to those of a more generaldisturbance , addressinga common difficulty in schizophreniaresearch . It is often argued that, in the presence of a general of in with e. degradation performance patients schizophrenia ( g., overall of it is difficult to know whether ), slowing response degradation in aparticular experimentalcondition is due to a specificdeficit or a more generalized one. However, this difficulty arisesprimarily when the mechanismsfor the deficits involved have not been specified . The model provides us with a tool for doing this. Above, we describedthe mechanismfor a specificattentional deficit related to disturbancesof dopamine activity in the PFC . To compare this to a more generalizeddeficit, we induced overall slowing in the model for each unit (cascade by decreasingthe rate at which information accumulates rate); this was done for all units in the model (figure 8.3C). A lower cascade rate induced an overall slowing of response , but no disproportionate , the specificdisturbancein slowing in the interferencecondition. In contrast context representationproduced both effects . Thus, the context hypothesis a better account for the data than at least one type of generalized provides deficit. We have explored others (e.g ., an increasein the responsethreshold), with similar results .

200

a J Tests andClinicalSyndromes Neuropsychologic

SIMULATION

OF THE CONTINUOUS

PERFORMANCE TFST

The Stroop model shows how contextual information and its attentional effects can be represented in a connectionist model , and how a specific biologically relevant disturbance in this mechanism can explain aspects of schizophrenic performance . One question we might ask is: How general are these findings ? Here , we extend the principles applied in the Stroop model to account for performance in the CPT . As we discussed earlier , patients with schizophrenia show consistent deficits in the CPT . This is particularly true for variants in which a demand is placed on memory for context . For example , in the CPT -double , a target consists of any consecutive reoccurrence of a letter (e.g ., aBimmediately following a B). Thus , subjects must remember the previous letter , which provides the necessary context for responding to the subsequent one. Patients with schizophrenia perform poorly in this task. This may be due to an impairment in the processing of context that , like deficits in the Stroop task, might be explained by a reduction of dopaminergic tone in the PFC. If this is so, then we should be able to simulate schizophrenic deficits in the CPT double using the same manipulation used to produce deficits in the Stroop task: a reduction of gain in the module responsible for representing and, in this case , maintaining , context . To test this , we constructed a network to the CPT -double . perform The network consisted of four modules : ( 1) an input module , (2) an intermediate (associative ) module , (3 ) a letter identification module , and (4 ) a response module (figure 8.4 ). The input module was used to represent the visual features of individual letters . Stimulus presentation was simulated by activating the input units corresponding to the features of the stimulus letter . The network was trained to associate these input patterns with the corresponding letter units in the letter identification module . In addition , the network was trained to activate the unit in the response module whenever a stimulus letter appeared twice or more in a row . This was made possible by introducing a set of connections from the letter units back to the intermediate units . This allowed the network to store and use information about the previous as well as the current stimulus (seeJ. D . Cohen & : Servan-Schreiber, 1992 , for a more complete account of training and processing in this model ). Note that there is a direct analogy between the role played by the letter units in this model , and the role played by the task demand units in the Stroop model . The representation over the letter units in the CPT model provided the context for disambiguating the response to a particular pattern of input , just as the task demand units did in the Stroop model . In the CPT model , however , context was determined by the previous input , and therefore changed from trial to trial . Following training , the network was able to perform the CPT -double task perfectly for a set of twenty - six different stimuli . To simulate the performance of normal subjects- who typically miss on 17 percent of trials and set off false alarms on 5 percent of trials (figure 8.S)- noise was added to the

201

-Schreiber& Servan : Cohen : Cognitive DeAcits in Schizophrenia

Letter Identification Module Response

t
Feature

Input

sesuodse

0.0 Nonnat = 1.0 ) Gain ( Schizophrenic (Gain = 0.61 )

in the for normalandsdtizophrenic andfalsealarms of misses subjects Figure8.5 Percentage and the simulation and for and circles test & Bed ( , continuous ) ( open squares squares perfonnance data . Empirical module circles ) run with normalandlow gainon unitsin the letteridentification -Kimling : Erlenmenger fromCornblatt ). aretaken ,& , Lenzenweger , (1989

202

Tests and Clinical Syndromes Neuropsychological

5 4

0 0

Empirical Data . MI . . False Alanna

Simulation Data DM ~ 0 Fal . . Alanna

net input, with the amount adjusted to match the performanceof the network with that of human subjects . The results of this simulation appear in = 6gure 8.S (gain 1.0). Then, to simulate the performanceof schizophrenic , we disturbed processingin the letter module ----which was responsible patients for representing and maintaining context- by decreasingthe gain of these units by an amount comparable to the amount used in the Stroop simulation (0.66). The percentageof missesincreasedto 46 percent , while false alarms increasedslightly to 8 percent . These numbers closely match the results of empirical observationsof schizophrenicsubjects . some authors have the Although interpreted performanceof patients with schizophreniain the CPT in terms of a de6cit in sensory processing , our model suggestsan alternative hypothesis : performancede6cits are due to a - as context- for degradation in the memory trace required processingthe current stimulus . We assumethat this memory trace is maintained in the PFC , and is directly influencedby changesin the dopaminergicsupply to this area . This hypothesis is consistentwith our account of Stroop performance , and with disturbances of languageprocessing , which we turn to next.

-DEPENDENT SIMULA nON OFCONTEXT LEXICAL DISAMBIGUA nON


The languagemodel (6gure 8.6) incorporates elementsof the two previous simulations . The network was similar to the CPT model. It was trained to associateinput patterns representing lexical stimuli (e.g ., the word PEN ) to : a responsemodule and a discoursemodule. patterns in two output modules Patterns in the responsemodule speci6edthe meaning of the input words " " (e.g., writing implement ), while the discoursemodule representedthe topic of the current sequenceof inputs (i.e., the meaning of the sentence , rather than the meaning of individual words). As in the CPT model, there were two -way connections between the intermediate module and the context (discourse ) module. Thus, once a discourserepresentationhad been activated an by input pattern, it could be used to influence the processing of subsequent stimuli in the semantic module. This provided the mechanismby which context could be usedto resolve lexical ambiguity . , the model was trained to produce local output Using backpropagation and discourserepresentations for thirty different input words, someof which were ambiguous . In the caseof ambiguouswords, the model was trained to produce the responseand discourse patterns related to one meaning (e.g., - + " writing implement PEN " and WRJ17NG - +, ) more than the other (e.g ., PEN " " fenced enclosure and FARMING . This ) asymmetry of training was similar to that of the Stroop model (trained on words more than colors), with acom parable result: when presentedwith an ambiguous input word , the network preferentially activated the strong (more &equently trained) response and discourserepresentations . To permit access to the weaker meaning , the network was sometimespresentedwith an ambiguousword as input along with

203

-Schreiber Servan & : Cohen : C~

tive Deficits in Schizophrenia

Meaning ( R esponse )Module . I Word Module Input

Disco C onte ( ) Modu

ISemantic ntermedi ( ) Module

-processingmodel. Patternsof activation over Figure 8.6 Schematicdiagram of the language the units in the input module are assumedto representthe current sensory stimulus (e.g., the orthographic code for a written word ), while the output module is assumedto represent the to generatean overt response(e.g ., the phonological code neededto pronounce information necessary the meaning of the word ). Note that the connections between the semanticand discourse modulesare bidirectional.

and FARMING one of its associateddiscourse representations(e.g ., PEN ), and " " trained to generate the appropriate response (i.e., fenced enclosure ). Finally, the network was trained on a set of context words, each of which was related to one meaning of an ambiguity; thesewords (e.g ., CHICKEN ) were " " trained to produce their own meaning as the response( fowl ) as well as a discourserepresentationthat was identical to the one for the corresponding ). The combined effects of these meaning of the related ambiguity (FARMING training procedureswas that when an ambiguous word was presentedand there was no representationactive over the discourseunits, the output was a blend of the two meanings of the word, with elements of the more frequently trained (dominant) meaning being more active than the other (subordinate . However, when a discourse representation was active, ) meaning the model success fully disambiguatedthe input and activated only the relevant . meaning response In this model, patterns of activation in the semanticmodule are derived from the learning algorithm rather than being determined by the experimenter . Eachpattern of activation produced by the activation of a particular " " unit ( word ) and of a discoursemodule unit is a distributed conjunctive input -within -FARMING ). representation of the word -within -discourse (e.g., PEN Such representations contain all the information necessary to activate a

204

Syndromes NeuropsychologicalTests and Clinical

single output unit corresponding to a given interpretation of the input (e.g ., " " fencedenclosure) on the responsemodule. - in very We tested the model' s ability to simulate simple form- the use of context in natural languageprocessing . Most words in English have more than one meaning ; languageprocessingrelies on context provided by prior stimuli to disambiguatecurrent ones . In the model, this occurred through activation of a discourse representation in response to each lexical input, which then provided context for processing subsequentstimuli. We tested the model for this ability by first presenting it with a word related to one of the meanings of an ambiguity (e.g ., CHICKEN), allowing activation to spreadthrough the network, then presenting the ambiguity (e.g ., PEN) and , the model was not directly observing the output. Note that, in this case . Rather , it had to construct this provided with a discourse representation from the first input, and then useit to disambiguatethe second . Tested in this with all context word word e. . either CHICKEN or way ambiguous pairs ( g , PAPERfollowed by PEN), the model was consistently able to generatethe meaningresponseappropriate for the context. To simulate performance in our experiment (see earlier discussion ), the model was presentedwith pairs of context and ambiguouswords (representing the clausesused in the experiment ) in either order. Following each pair, ' the network was probed with the ambiguousword , simulating the subjects processof reminding themselvesof the ambiguity, and choosing its meaning . At each time step of processing , a small amount of noise was added to the activation of every unit. The amount of noise was adjusted so that the simulation produced an overall error rate comparableto that of control subjects . The model' s responseon each trial was consideredto be the meaning that was most active over the output units after the probe was presented . To simulate the performance of patients with schizophrenia , we introduced a disturbanceanalogousto the one in the CPT model: a reduction in gain of units in the context module. The results of this simulation (shown in figure 8.78) show a strong resemblance to the empirical data (figure 8.7A ). They demonstrateboth significant effects : (a) in the low gain mode, the simulation makesabout as many more dominant responseerrors as do the subjectswith schizophrenia ; however, (b) as with the human subjects , this only occurred when context camefirst. The model provides a clear view of this relationship between dominant response bias and memory. When gain is reduced in the context module, the representation of context is degraded ; as a consequence , it is more susceptibleto the cumulative effectsof noise. If a contextual representationis used quickly, theseeffectsare less , and the representationis sufficient to overcome adom inant responsebias. However, if time passes(as when context is presented first), the effects of noise accumulate , and the context representation is no longer strong enough to mediate the weaker of the two competing . responses

205

-Schreiber Servan & : Cohen : Cognitive Deficits in Schizophrenia

SJOJJ8

0 .

Schizophrenics Patient contois

0 .

Low gain (In context module ) Nonnal gain

in two of the threeconditions run in the empirical study Figure 8.7 Errorratesfor subjects and for the language model simulation . The rate of unrelated erron in all conditions(not shown whenthe strongmeaning wascorrect ), andof weakmeaning (alsonot shown ) responses wereapproximately the same in the normalandlow-gain conditions of the simulation . andof the same asthatobserved in human (about1 to 2 percent ). subjects magnitude

CONCLUSION
The three models we have presentedshowed how the connectionist framework . This was can be used to link biological and behavioral observations achieved by bringing a common set of mechanismsto bear simultaneously on physiological and psychological phenomena . Speci6cally , the models : (a) they simulated quantitative aspectsof performance served several purposes in three previously unrelated behavioral tasks ; (b) they elucidated the role of processingof context in both the attentional and linguistic tasks ; (c) related of context to es and d ; ( ) they they biological process processing showed how a specific disturbanceat the biological level could account for . The models , and the simulations schizophrenic patterns of performance based on them, relied on many simplifying assumptionsand provided, at best, a coarseapproximation of the mechanisms underlying both normal and for behavior. While accounting empirical data is a primary schizophrenic , McClelland (1988) goal in the development of computer simulation models has argued that this may not be the only basis for their evaluation. Models are useful if, among other things, they offer new interpretations of empirical . We have indicated phenomenaand unify previously unrelated observations how our models simple as they are - may fulfill these different functions. Finally, models can be used to derive predictions that can be tested in . We have recently found connnnation for predictions made empirical studies

206

Tests andClinical Syndromes Neuropsychological

first

Context

last

Context

first

Context

last

Context

0 0 . . . . .

. 0 2 8 6 4

12 10

by our models using a variant of the CPT task described here (Servan Schreiber Cohen & : , , Steingard , 1996). Although our models may have met thesespecificcriteria, our overarching hope is that they will have helped provide a more refined and integrated approach to the riddle of behavioral and biological disturbancesin schizophrenia and to the study of brain-behavior relationshipsin general .

ACKNOWLEDGMENT
This work was supported by an NIMH Physician Scientist Award D. Cohenand an NIMH IndividualFellowAward ( MHOO673 ) to Johathan -Schreiber . ( MHO9696 ) to DavidServan

REFERENCES
" " . R., JordanD. E., & : Hegel .R , M. (1983 Abramczyk ). ReverseStroopeffectin the performance of schizophrenics . Perceptual andMotorSkills . , 56 , 99- 106 Brozoski . M., Rosvold : Goldman , T. J., Brown , R , H. E., & , P. S. (1979 ). Cognitivedefidt caused of dopamine in prefrontal cortexof Rhesus . Science by regionaldepletion , 205 , 929 monkey 931 . . P., & : Miller, G. A. (1964 , L J., Chapman ,J in schizophrenia ). A theoryof verbalbehavior Chapman . In B. A. Maher .), Progress in uptrimentaltit T , research (Ed S On Rlihi . Vol 1. New York : AcademicPress .

- induced Chiodo : Berger , L. A , & ). Interadions between dopamine andaminoacid --- , T. W. (1986 . excitation andinhibitionin the striatum . BminResearch - 203 . , 375 , 198 Cohen : Mcdelland , J. D., Dunbar , K., & , J. L. (1990 es :A ). On the controlof automatic process - 361 distributed modelof theStroopeffect . Psychological Review . , 97 parallel (3), 332 processing Cohen . D. & : Servan Schreiber D. ,J , , : A connectionist (1992 ). Context , cortexand dopamine to behavior andbiologyin schizophrenia . Psychological Review . , 99 approach ( 1), 45- 77 -Schreiber Cohen , J. D., Servan : Spiegel , D., Targ , D. (1992 : ). Thefabricof thoughtdisorder , E., & A cognitive neuroscience to disturbances in theprocessing of contextin schizophrenia . approach In D. Stein : J. Young(Eds .), Cognitive science , & andclinical disorders : (pp. 99- 126 ). New York Academic Press . Cohen . M., Semple ,R , W. E., Gross , M., Nordahl , T. E., DeUsi , L. E., Holcomb , H. H., King ,A C., Morihisa : Pickar in a prefrontal , J. M., & , D. (1987 substrate of sustained ). Dysfunction attention - 2039 in schizophrenia . LifeSciences . , 40 , 2031 -Kimiing Comblatt : Erlenmeyer , B., Lenzenweger , M. F., & ). A continuous , L. (1989 performance test : II. Contrasting , identical attentional and depressed pairsversion profilesin schizophrenic . Psychiatry Research , 29 , 65- 85. patients Diamond and neural , A (1989 bases ). Thedevelopment . New York : of higher cognitive functions NewYork Academy of Science . Foote : Oliver , S. L., Freedman , R., & of putativeneurotransmitters , A P. (1975 on neuronal ). Effects . BminResearch - 242 . activityin monkey , 86 , 229 auditorycortex Freeman , W. J. (1979 corticalstimulus ). Nonlinear -response relations . Biological gainmediating - 247 . , 33 , 243 Cybernetics

207

-Schreiber Servan & : Cohen : CQ2nitive Deficits in Schizophrenia

Fuster , J. M . (1997). Theprefrontalcoria : Anatomy , physiologyand neuropsychology of the frontal lobe . , 3rd ed. New York: RavenPress -Rakic Goldman , P. S. (1987). Circuitry of primate prefrontal cortex and regulation of behavior .), Handbook : Thenervous by representational , Vol . 5 memory. In F. Plum (Ed of physiology system (pp. 373- 417). Bethesda , MD : American PsysiologicalSociety. : Franzen , G. (1974). Abnormalities of cerebralflow distribution in patients with Ingvar, D. H., & chronic schizophrenia . Acta Psychiatric a Scandinavia , 50, 425- 462. of drug effectsin schizophrenic Kometsky, C. (1972). The use of a simple test of attention as a measure . Psychophamracology , (Berlin ), 8, 99- 106. patients Malmo, H. P. (1974). On frontal lobe functions: Psychiatric patient controls. Cortex , 10 , 231237. McClelland, J. L. (1988). Connectionist models and psychological evidence . Journalof Memory and Language , 27, 107- 123. Meltzer, H. Y., & : Stahl , S. M . (1976). The dopamine hypothesis of schizophrenia : A review. BuiltHn , 2, 19- 76. Schizophrenia Nuechterlein , K. H. (1984). Information processingand attentional functioning in the developmental courseof schizophrenicdisorders . Schizophrenia Bulle Hn , 10 , 160- 203. -Schreiber Servan : Steingard , D., Cohen , J. D., & , S. (1996). Sdtizophrenic deficits in the processing of context: A test of neural network simulations of cognitive functioning in schizophrenia . Archivesof Gmeral Psychiatry , 53, 1105- 1112. -Schreiber Servan : Cohen , D., Printz, H., & , J. D. (1990). A network model of catecholamine effects : Gain , signal-to- noise ratio and behavior. Science , 249 , 892- 895. : Krus, D. M . (1960). Effectsof lysergic acid diethylamide , S., & , and differencesbetween Wapner normalsand schizophrenics on the Stroop color-word test. }oumRlof Neuropsychiatry , 2, 76- 81. : Berman . F. (1988). Speculation on the meaning -, D. R., & of metabolic hypo, K Weinberger . Schizophrenia Bulletin , 14 , 151- 168. frontality in schizophrenia . F., & : Zec, R . F. (1986). Physiological dysf Wtctionof dorsolateral , D. R., Berman , K Weinberger cortex in schizophrenia : I. Regional cerebralblood flow evidence . Archivesof General prefrontal

- 125 . , 43 , 114 Psychi Rtry

: Sweet of brain , J. J., & - damaged , J. J. (1985 ). Identification , schizophrenic Wysocki , andnormal medical . Int~ tional -l of dini patients usinga briefneuropsychological }ourna battery screening CRI ,7 (1), 40- 49. Neuropsychology

208

Tests andClinicalSyndromes Neuropsychological

Neural Network Modeling of Executive Functioning with the Tower of Hanoi Test in Frontal Lobe - Lesioned Patients andRandolph W. Parks JohnCardoso

Some of the common elementsof executive functioning in humans include abstract reasoning , problem solving, mental flexibility , responseinhibition , self-monitoring, and maintenance of mental sets (Fletcher , 1996; Walsh, 1978). Lezak(1995) suggestedthat a key elementin executive functioning is " planning: The identification and organization of the steps and elements needed to carry out an intention or achieve a goal constitute planning . . ." (pp. 653- 654). Thesedefinitions, while not exhaustive , constitute the critical elements in successfulcompletion of the Tower of Hanoi (TOH ) neuropsychological test. In contrast , brain- damagedpersons with anterior brain lesions often perseverateresponsesand exhibit failure to maintain set on tests such as verbal fluency and the Stroop (Benton, 1968; Crockett, Bilsker , Hurwitz , et al., 1986; Hoist & Vilkki , 1988; Perret , 1974), and Wisconsin Card Sorting tests (Arnett , Rao in, et al., 1994; Robinson , Bernard , Heaton, Lehman , et al., 1980). Other researchhas not supported the emphasis of these tests to anterior brain regions, noting that size , shape , depth, and critical impact of damaged neural circuitry are also very importantS . W. Anderson, Damasio , Jones , et al., 1991; Dodrill , 1997; Newcombe , 1969). Sinceits introduction, apparently by de Parville in La Nature (Paris , 1883), the TOH puzzle has challengedthe problem-solving and planning abilities of . Mathematicians , computer scientists many researchers , and artificial intelligence researchers view the TOH test as an important example of recursion , one of the fundamental techniques used in the design of computer algorithms . Many computer sciencetextbooks use the TOH to teach recursion . While the TOH can easily be solved by a computer , using either recursiveor iterative algorithms and traditional procedural programming methods , we utilized artificial neural networks to solve the task.

CLINICAL NEUROPSYCHOLOGICAL DATA


For many years the TOH and related versions have been used in clinical assessment of various frontal populations (Owen, Downes, Sahakian , et al., 1990). Shallice (1982), for example , conducted a study of brain-lesioned groups (left frontal, left anterior, and homologous right -lesionedpatients) to

control subjectson a related test (Tower of London). He found the left anterior group was most impaired in a timed version of the test comparedwith the other groups. Glosser and Goodglass (1990), in a similar experiment with the TOH (using moves and violations as variables ), found only the left frontal group was impaired comparedwith the other lesion groups and controls . In our study, we focusedon those neural network attributes concerned primarily with the left frontal lobe. This was done becauseof frontal lobe heterogeneity and becausemost of the above-named studies that supported a left frontal primacy. To a lesserextent, we a frontal hypothesis emphasized added neural network parameters concerned with attentional factors and , since a brain imaging study had also implicated visuospatial components these latter regions (Andreasen , , Alliget , et al., 1992). In this chapter , Rezai how the we shall demonstrate networks , computer using backpropagation learned to solve the TOH test. In addition, we examined the effects of " " " " deliberately harming (computer programmers use the terms mutilation " " or degraded to refer to these large changesin synaptic connections ) the neural network by massively altering the values of large sections of the . net' s weights with data reflecting left frontal lobe lesions and Goodglass were from Glosser The neuropsychologicaldata acquired of frontal executive (1990). Briefly, those authors used several test measures and in various left right hemispherenonaphasic functioning hemisphereaphasic with normal controls. The part and them damagedpatients compared that the left frontal was the of the experiment relevant to ours finding 's group performancewas significantly impaired compared( MANOV A [multiple : combined moves and violations variate analysis of variance ] calculations . The mean number of TOH moves on the TOH with normal controls ) = = 4.44 in the left frontal group M 14.00 was SO number of moves , ) (total = = and M 10.13, SO 3.62 in the normal control group. STRATEGIFS TO SOLVE THE TOWER OF HANOI PUZZLE Solving the TOH (see figure 1.2 in chapter 1) consists in moving the rings from their original peg (the sourcepeg) to one of the other pegs (the target peg), and stacking them in the sameinitial order by using two simple rules: (1) move only one ring at a time and (2) never stack a ring on top of a smaller one. For n number of rings, the minimum number of moves capable of solving the puzzle is 2" - 1. For instance , for a three-ring tower it will take at least sevenmoves to solve the puzzle. The apparentsimplicity of the . Although a three-ring puzzle is relatively TOH puzzle can be misleading , once the number of rings is easy to solve in the least number of steps increased , the optimal solution to the puzzle becomesharder to find. Careful and observation of people engaged in solving the TOH show that testing . there are different strategiesthat can be used " " of In his paper The FunctionalEquivalence Problem Solving Skills, H. A . Simon (1975) examined four of those types of strategies , respectively: (1)

210

andClinical Tests Neuropsychological Syndromes

rote , (2) goal -recursion , (3 ) perceptual , and (4) move -pattern strategies . Each of these was considered to have its psychological consequences; for instance, different strategies place different loads on short - term memory . Simon attempted to address what was involved in learning to solve the TOH puzzle . A computer program was created with the purpose of clearly identifying the steps and strategies that must be learned to solve the TOH problem . The General Problem Solver (GPS), a program that uses meansends analysis as its primary method , and which appeared, in the opinion of the authors ( Newell & : Simon , 1972 ), to describe well the organization of human problem -solving processes, was also used. The GPS program has been success fully used to solve problems in which each situation , or step, can be compared with a goal situation , to discover one or more differences between them . The program uses a set of operators that can be used to eliminate the differences between the current and goal situations . If an operator cannot be applied , the current situation is compared with the conditions for applying the operator , and if some difference is discovered between them , the GPS is then applied recursively to eliminate this difference . Simon ' s conclusion was that , except for the rote programs , the programs used to solve the puzzle all used the same type of means-ends analysis used by the GPS program . How would one solve the TOH puzzle in the early 1970s1 Starting with the simplest possible case , a tower of one ring ( n = 1) , we simply transfer the ring from the source peg to the target peg and complete the puzzle in the minimum possible number of moves , that is, 2" - 1 or one single move . The degree of difficulty increases with a two - ring tower ( n = 2 ) . To solve a two -ring tower we move the small ring to the spare peg and move the large ring to the target peg . Then we move the small ring from the spare peg to the target peg , and stack it on the top of the large ring, thus completing the " puzzle in the minimum possible number of moves , that is, 2 - 1, or three moves . What we need to do is to store the small ring momenta rily on the spare peg to allow for the move of the larger ring from the source peg to the target peg . " Finally , for a three - ring tower ( n = 3 ) : according to our formula ( 2 - 1), we should be able to solve the puzzle in seven moves . The following lists the moves : 1. Move the small ring to the empty target peg . 2. Move the middle ring to the empty spare peg . 3. Move the small ring to the spare peg .

4. Move the large ring to the empty target peg. S. Move the small ring to the empty sourcepeg. 6. Move the middle ring to the target peg. 7. Move the small ring to the target peg.

211

Cardoso& I: Parks : ExecutiveFunctioning

Note that we had to make room to move the large ring to the target peg. Once the large ring was moved from the source peg to the target peg, the , and one that we have already solved, the problem problem becamesimpler of moving a two -ring tower. In summary , moving a tower of n rings from the source peg to the target peg is equivalent to moving a tower of n - 1 rings to the sparepeg , the largest ring from the sourceto the target, and the n - 1 ring tower from the spareto the target. We can use a recursiveprocedure to solve the puzzle .A , becausethe problem is recurrent in character Tower recursive procedure to solve the TOH puzzle can be expressedin as follows: pseudocode Tower ( n , source , target IF n = l THEN Move ( n , source , target ) ; ELSE Tower ( n - l , source , spare ) , Move ( n , source , target ) , Tower ( n - l , spare , source ) . This recursive procedure will succeedin solving the TOH puzzle for three pegs and for n rings. Although recursive procedures are very powerful mechanismsthat provide simple and elegant ways of thinking about these , they do not always use computer resources , such as time types of problems and working memory, in the most efficient manner . Problemsof a recurrent nature can also be solved using iterative procedures . Iterative proceduresuse in little with their recursive . There is, very memory comparison counterparts however, a price to be paid: the program has to keep track of the position of the rings at all times. Whether a computer program uses recursion or iteration to solve the TOH puzzle , the programmer will have to provide the computer with a set of commandsthat must be followed in a predetermined sequenceto reach the correct solution. Computer programs are usually rigid and inflexible. Since it is doubtful that the human brain uses these types of operations to solve everyday problems , researchershave looked into other problem, such as artificial neural networks, that could bepotentially solving techniques closer to the brain' s methods and help to understandhow neuronsmay learn and operate . ):

ENCODINGTHE MOVFS
The TOH , like its counterparts Tower of London (hierarchically arranged ) and Tower of Toronto (adding a fourth ring), requiresthe ability training sets to anticipate subroutine steps to achievethe desired final position of the

212

NeuropsychologicalTests and Clinical Syndromes

rings. To construct our current network we had to conduct some theoretical mathematicaltests to assistwith the development of the details of the computer architecture , such as learning rule, placement of rings, and efficient , 1991). There are many different types of sequencingparameters(Cardoso neural nets and different types of training algorithms to choose from (Lippmann forward , multilayer architecture using the , 1987). The choice of a feed backpropagationtraining algorithm to solve the TOH was made becauseof : Rumelhart , and reliability (Chauvin & versatility, learning capabilities , 1995). The results are describedbelow, after considering some general information about the computational ability and the functionality of neural nets. How can we represent , or encode , the TOH puzzle so that the net learns and later remembers how to solve the puzzle ? The graphical display of 9.1 in the . figure helps understanding encoding method that was selected " The Initial Position" diagram depicts the three disks , stacked on the leftmost peg by order of size with the largest disk on the bottom of the pile (denoted by the three horizontally adjacent black squares ) and the smallest

Empty Pegs

Initial Position

15t move

2nd move

7th move
used to encodethe positions of the Towerof Hanoipuzzle . Figure9.1 Representation

213

Cardoso & : Parks : ExecutiveFunctioning

disk on top (denoted by the single black squareon the topmost row ). The following sections of the picture (1st to 7th moves) show the shortest of moves necessary to solve the three-disk TOH puzzle. Note that sequence we need nine small squaresto represent the condition of each peg at any to representthe state of the given time. We need twenty -sevensmall squares puzzle. Eachone of the small squareswill be encodedas a separateinput to the net. We proceed &om this intermediate graphical representationto the final mathematicalvector representationby encoding a vacant (white) square as a 0 (zero) and an occupied (black ) squareas a 1 (one), while moving &om left to right on each row , and &om the top to the bottom row. With this method, the initial position of the puzzle is representedby the vector ( 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0), and the final position by the vector (000000 100000000 110000000 111 ). Both vectors have a length of twenty -seven . In addition, the network will need some sort of " context inputs." Three of , are every nine squaresof eachpeg, depicted in figure 9.1 as shadedsquares reserved to be used as context inputs. These context inputs are necessary because , in most cases , for every position in the puzzle there are usually three legal moves. During the training phase , we have to provide a mechanism to help the neural net to differentiate these moves. For this reason , associatedwith each move there is a " move-quality descriptor" ( M-Q). There are three types of M -Q descriptors , one for each move type, respec " " " " ' " , and long . tively : best, regressive The number of M -Q descriptors is unrelated to the number of context squaresper peg; the number of context squares per peg was arbitrarily . Best moves (moves that lead to the shortest route) are indicated by chosen the presenceof one or more (one, two , or three) context squaresin the left peg. Long moves (moves that lead to a longer route) are encoded by the presence of one or more context squares in the middle peg. Regressive moves (moves that take you back to some previously visited position) are indicated by the presenceof one or more context squaresin the right peg. In this experiment , the M -Q descriptorsuse all three context inputs.

THE TRAINING SET SELECTING


There were thirty -nine input patterns in the training set, eachpattern representing a given position in the progression to solve the tower puzzle. The selection and the numbering of the patterns can be described as a directed graph (figure 9.2). The patterns used are only a small subset of all possible . The nodes (boxes ) in the graph of figure 9.2 stand for a given state patterns or of the TOH . The numbers) inside eachbox serve only ( position) puzzle to identify a particular state . The directed arrows leaving a node (marked '1" " " for long, " r" for regressive , and b for best) represent moves, and clearly show what states are possible to reach &om that node. The node numbers are only an identification (10) tag; they do not have any other specialmean

214

NeuropsychologicalTests and Clinical Syndromes

. @ . : ~ ~ ! . # ; @ : ~ r . . 1 ~ ! @ # ; ' b ~ t r , b . # ; ' " ~ : @ 1 # ; '@ b ~ ~ t r y . . r r 1 9 2 3 " , # ; ' : 7 ~ ~ _ . @ # ; '1 0 ~ ~ \ r : !( 1 ~ ~ t .b


21S Cardoso & : Parks : Executive Functioning

of the succession of moves selected to train the neural net to solve Figure9.2 Representation theTowerof Hanoipuzzle .

ing nor are they usedby the net at any time. They are used only to facilitate our identification of the statesor positions of the puzzle. The large ID numbers that appear in the leftmost column of nodes (312, 302, 203, and 213), unlike the others, are not a result of any numerical sequence ; they were arbichosen to the and size of the trarily represent position rings that sit at the bottom of each peg. Seefigure 9.3 to understandhow these numbers were chosen . Figure 9.3 is another way of representing what positions can be reached during the first two moves of a three-ring puzzle. Eachlarge squarehas three columns and representsthe state of the puzzle at any given time. Eachcolumn , of three small squares , representsa different peg. The digits within the squaresrepresent the rings, with the largest digit representing the largest " " " " ring . A zero stands for empty. The digits precededby # , and located outside the large squares , are the ID tags of the states , or positions, of the puzzle. The first move from a three-ring puzzle can take us to two different states or positions, respectively : states # 2 and # 3. In this example we will not consider the sequence of moves derived from state # 2; we will use only the of moves originated from state # 3. From state # 3 there are a total sequence of four possible reachablestates : #4, # 5, #6, and # 7. We will , respectively discardstate # 5 because it is an illegal state (no ring can be placed on top of a smallerone). The secondmove can thus take us to one of three valid positions : #4 (the best move), #6 (a regressivemove), and # 7 (a long move).

oo OO 200 200 310 301 ~ il E ~ ~ ~ : : ~ # 40 1 ? ~ ~::;: 0 0 0000 2200200 321 300310 1E0 iimJJ :0 ' 1E0
of various Tower of Hanoi moves and illegal state (crossedout ). Figure 9.3 Representation

oo 0 0 2 3 0 0 ~; " ' / " 1L~ 3 ~

This brief explanation should elucidate how the positions were identified and the moves . Some of the boxes in figure 9.2 do contain more than one number because some states appear to repeat . For instance, in figure 9.3 we can see that state # 2 is identical to state # 7, and state # 1 is identical to state # 6 . The shortest sequence of moves to solve the three -ring TOH . puzzle is shown , in figure 9.2, in the rightmost column of nodes. As we move to the columns on the left we find ourselves taking longer routes to the solution . One can think of each line of the training set as representing the arrows between nodes of the graph in figure 9.3. There are thirty -nine lines in the training - set table. The reader is cautioned that while each of the thirty -nine lines might be interpreted as " moving a ring in the TOH ," this is not necessarily the case. One of the functions of the thirty - nine lines of input / output vectors is their use as placeholders in a computer matrix . In their function as placeholders , the network is allowed to choose from a range of responses during training . The more a neural network is degraded , or , to use a human metaphor , lesioned , the number of input / output vectors being utilized is likely to increase. Each line has two vectors : the first vector (on the left ) is the input pattern , the second vector is the target pattern (the expected output from the net for its corresponding input pattern ). In supervised learning during each training step we need to provide the net with both vectors simultaneously . Here the network must learn to associate the input vector with the target vector . There are thirty -nine input -target vector pairs (see figure 9.4). As mentioned before , the move from state # 1 to state # 2 was arbitrarily disregarded because it leads to a different sequence of moves (one that ends with the rings stacked on a different peg, i.e., the center peg ). The net was trained to generate the target patterns when it was presented with the input patterns . For example , in the first sample of the training , the

216

NeuropsychologicalTests and Clinical Syndromes

Input vector 111000000111000000111000000 011000000111000000111000100 000000011110000001111000100 000011000110001000111000100 000000011000000001111110100 000011000000001000111110100 011000000001000000111110100 000000011110000001111100000 011000000111000000111100000 000011000110001000111100000 000000011000000001111100110 011000000001000000111100110 011000000101000000111000110 000000011100000001111000110 011000000101000000111110000 000000011100000001111110000 000011000100001000111110000 011000000001100000111110000 000011000000101000111110000 000000011000100001111110000 011000000001100000000110111 000011000000101000000110111 000000011000100001000110111 000000011000000001100110111 000011000000001000100110111 011000000001000000100110111 011000000001000100000110111 000000011000000101000110111 000011000000001100000110111 000000011000000101110000111 011000000001000100110000111 000000011000000001110100111 011000000001000000110100111 000011000000001110100000111 011000000001000110100000111 000000011000000111100000111 011000000001000110000100111 000000011000000111000100111 000011000000001110000100111

Targetvector 000000000110000000111000100 000000000000000000111110100 100000000110000000111000000 000000000110000000111100000 000000000110000000111000100 000000000100000000111110000 000000000000100000111110000 100000000110000000111000000 000000000110000000111000100 000000000000000000111100110 000000000110000000111100000 000000000100000000111000110 000000000100000000111110000 000000000000000000111100110 000000000000100000111110000 000000000000000000111110100 000000000100000000111000110 000000000000100000000110111 000000000100000000111110000 ~ 00111110100 000000000000000 000000000000000000100110111 000000000000000100000110111 000000000000100000111110000 000000000000100000000110111 000000000000000100000110111 000000000000000110100000111 000000000000000000100110111 000000000000100000000110111 000000000000000100110000111 000000000000000100000110111 000000000000000000110100111 000000000000000100110000111 000000000000000110000100111 000000000000000110000100111 000000100000000110000000111 000000000000000000100110111 000000100000000110000000111 000000000000000110100000111 000000000000000000110100111

Figure 9.4

of input and target pairs. Representations

input-pattern vector is: (1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0 1 1 1 0 0 0 0 0 0). When presented with this pattern, the net was trained to generate the target-pattern vector: (0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 1 1 0 0 0 1 0 0). Note that included in each input-pattern there are also the bits that encode the M -Q descriptor for that move. In the example above, for easy identification , moving from position , theseare indicated in bold type. In this case " " # 1 to position # 3 is encodedas the best move.

217

t Parks : ExecutiveFunctioning Cardoso&

THENEURAL TRAINING NE1W 0 RK


The backpropagationalgorithm, basedupon the generalizeddelta rule, was selectedto train the feed forward multilayer network. This learning algorithm hasbeendiscussed : Rumelhart , 1995; Pao , 1989) by other authors (Chauvin & and the mathematicsof the processwill not be reviewed here. As mentioned in chapter 1, the input is propagated forward, from the input layer through the middle layer to the output layer, and then the resulting output is propagated backward , from the output layer through the middle to the input layer, to compute the error 15 . Subsequently , the weight error derivatives and the are . The network was trained using the parameters weight changes computed shown in table 9.1. Training the network consisted of repeatedly presenting the patterns in the training set to the net until the generatederror was acceptable . The patterns were randomly selectedfor presentation , and a random value between :f: 9.9 percent of noise was added to each input. The graph of figure 9.5 shows the learning curve of the network. The sharpinitial slope of the curve and its slow settling inmcatesthat the net had no difficulty in learning the thirty -nine input patterns in the training. Learning was completed in 431 iterations. That is, the patterns in the training set were preTable9.1 Parameters used during Tower of Hanoi learning
Parameter Type Momentum Learningrate Maximum total error Maximum individual-unit error Value

0.9 0.7 0.01 0.001

~ ~ III
9 4

Figure 9.5 Learning curve for the 27- 15-25 neural network during the Tower of Hanoi train. in~ session .

218

Tests andClinicalSyndromes Neuropsychological

14

Learning Iterations

sented to the net 431 times before the net could learn to associate each input pattern with its corresponding output pattern . Clearly this number could have been reduced if larger error values had been selected; in our case , this was not a concern.

THE EFFECTS OFLEARNING


The purpose of the learning pha~ e is to compute a set of weights that will allow the network to execute its job adequately . In our experiment , this refers to one of the TO H' s positions and the ability of the network to produce one of the next allowable positions. In a three-layer network there are -layerweights(the two distinct weight sets : the hidden , respectively weights of the connections between the input and the hidden layers) and the output layer weights(the weights of the connectionsbetween the hidden and output ). The hidden-layer weights are understood to be responsible for layers allowing the hidden-layer neuronsto extract relevant features&om the input . These featuresare then passedto the output neurons patterns , through the , and the net will then generateits final output . It is output-layer connections generally understood that for a successfulknowledge distribution over all -zero values . Theseweights will have a weights, most weights will have near -normal distribution around zero. Other near weight distributions are generally a sign of local representation , rather than distributed. A distributed is the most desirable . As in the brain, a distributed representation generally will the &om representation system protect forgetting all it has learned in case of injury or malfunctioning. In such cases , whether by accidental or ' deliberatedamage , the systems performancewill suffer little or no degradation , and the systemwill not be renderedtotally useless . 9.6 shows the of the &equency distribution of the Figure histogram -zero values hidden-layer weights; most of these weights have near . The kurtosis (curvedness ) of the distribution is 3.638, indicating that the distribution is fairly sharp . The skewness(symmetry measure ) of the distribution is

Bin Values

-layerweights distribution of hidden . Figure9.6 Frequency

219

Cardoso& : Parks : ExecutiveFunctioning

160 140 120 ~ c 100 l 80 ! 0 " 60 ! & I. 40 20 0- - - - - - - 987-6543210123456789 Bn Valu .. 9 .7Frequency distribution of . layer Figure output weights
Input Activity from Tr8nlng Set

~ 25 20 15 10 5 0

11 13 15 17 19 21 23 25 27 Inputs

Figure 9.8 Activity generatedin the input layer by the patterns in the training set and its lin@ ar tr~ lin@ -

1.930, indicating that the scorestrail off to the right (positively skewed ). The histogram in figure 9.7 shows the frequency distribution of the output -layer , with a kurtosis value of 9.139, and the weights. The curve is much sharper skewness , 2.985, is still positive (trailing to the right ) but larger. Let us look now at the distribution of the input activity during training. " " , is the total number of active signals that reach a Activity , in our case given input . An input is said to be active when the incoming signal is a 1 (one). Figure 9.8 graphs the activity distribution over the input lines. Over the thirty -nine input patterns in the training set there is a total of 1053 . This disparity may , that is, 351 ones and 702 zeros (27 x 39) input signals explain the inhibitory trend of the network weights. The graph of figure 9.8 shows most of the activity located over the last third of the inputs (between 19 and 27). This is a direct consequence of the type of encoding that was " chosen(seefigure 9.1). This active area correspondsto the inputs that " see the bottom layer of the pegs , which is always the busiest of the three layers. The line acrossthe graph plots the linear trend of the data.

220

Tests andClinical Syndromes Neuropsychological

Table 9.2 Best Moves 931 (0.26)

Resultsof testing the unmutilated network with the set of 3584 patterns Long Moves 666 (0.19) Moves Regressive 292 (0.08) Other Moves 1695 (0.47)

: Numbersin parentheses Note representpercentageratios.

TFSTINGTHE NEURALNETWORK
After learning , the neural net was tested with a set of all 3584 (512 x 7) possiblepattern variations derived &om positions 1, 3, 4, 13, 15, 18, and 25. These positions are those that solve the puzzle in the minimum number of moves (see rightmost column of figure 9.2). Pattern variations, for a given ' position, are obtained by OR-ing the position s true pattern (the bits that represent possible ring positions) with all possible binary combinations of the nine M -Q descriptor bits. For instance , the true pattern for position 1 is lx Oxllx O Ox O Ox I II000000 ), where the symbols " x " are placeholdersfor ( Ox the M -Q descriptors . There are nine M -Q descriptor bits in each pattern. Sinceeachof thesebits can have one of two values(0 or 1), we can create29 (512) variations of a given pattern for a total of 3584 for all seven chosen . These3584 patterns constituted the testing set that was used in our patterns tests. The first tests that modeled the normal controls used the network with all its connedions intact (nondegraded ). The results are summarizedin table " 9.2. Correct responses include "best," 'long ," and " regressive moves. Incorrect " " the in the Moves Other column include all other , , figure responses = network responses( 1695 3584 931 666 292); these responsesmay include false moves , nonexistent pattern codes , or existent but unreachable or invalid moves. In this test, the best best + long+ regressive /( ) move ratio is 0.5 931 931 666 . Let us call this ratio the performance 292 + + ( /( approximately index value (PlY) of the net. Assuming that we take the PIV of 0.5 as the measureof the performanceof our " normal" (unmutilated) net in solving " " ' the puzzle , we can then contrast this value with a normal persons PIV. An experimental study (Glosser & : Goodglass , 1990) has shown that a " normal" , on average , ten moves to solve the TOH puzzle , person takes while a left &ontal lobe- damagedperson takes fourteen moves. Since the minimum possible number of moves is seven , we can assumethat a normal ' PlY s is 0.7. person Although , by simply looking at those approximately , we do not know the types of the moves each person makes , this figures value arises&om the fact that if one takes more than seven moves to solve ' " the puzzle , then some of these moves must be &om the long and/ or " " regressive type. DEGRADADON OF THE NEURAL NETWORK Recall that there are twenty -seven output neurons . Eachoutput neuron has sixteen input connections , each one of these connectionslinking the neuron

221

Cardoso & : Parks : Executive Functioning

to eachneuron in the hidden layer; there is one additional connection from a bias neuron. We can store the weights of theseconnectionsin a 27 row x 16 column weight matrix ( W ). The rows correspondto the output neurons and the columns correspond to the hidden neurons (i.e., the value in W [23, 10] standsfor the weight of the connection between hidden neuron 10 and output neuron 23). To investigate the results of degradation of different sectors of the output-layer connections on the performance of the network, this , corresponding to the weight matrix was subdivided into eight submatrices - 27, 1- 8 , and 1 14 1 left 1 13 16 bottom 27 , , , ([ W1 (W [ 161 [ top ( W 1 l halves and to the left 13 , 1- 8 , top-right (W [ ( 27, 9 16 right ( W top [13- 27, (W [ 1 13, 9 16 , bottom left ( W [ 13 27, 1 8 , and bottom -right ( W . Degradation of a weight 9 16 quarters of the output layer connections submatrix consisted in reducing its values to 1 percent of its initial value. This drastic change forced the values of these weights to near zero, practically turning these connections off. The network, with its degraded connections . , was then tested using the same 3584 patterns previously used submatrix. The one for each tests were , degraded Eight separate performed results of thesetests are displayed in table 9.3. A left frontal lobe- damagedpatient takes , fourteen moves to , on average solve the puzzle. Using the same criteria that we used before, we can also compute the PlY for the averagefrontal lobe- damagedperson as the ratio best + long+ regressive , which would be approximately 0.5 ) moves / (best , if a nongradednetwork, with a PlY of (7/ 14). Assuming a linear relationship a normal 0.5 to , with a PIV of 0.7, then a , person approximately corresponds would with a PlY around 0.35 net , , correspond to a left frontal degraded lobe- damaged patient with a PlY of 0.5. The network with a degraded bottom-right quarter, similarly to a frontal lobe- damagedpatient, has a PlY of approximately 0.35 (table 9.4). To investigate the mutilation effects on the responsesto a single move, one pattern (pattern for position # 3) was randomly chosen and tested first

222

andClinical Syndromes NeuropsychologicalTests

Table 9.4 Degraded Top half Bottom half Left half Right half

perfonnanceindex values(P I Ys ) for the various degradednetworks PlY (approx.) 0.50 0.50 0.70 0.00 0.58 0.02 0.66 0.36

Top- left quarter Top- right quarter Bottom-left quarter Bottom-right quarter

Table 9.5 Resultsof testing an nondegradednetwork using position # 3 patterns Best Moves 35 (0.07) Long Moves 273 (0.53) Moves Regressive 44 (0.09) Other Moves 160 (0.31)

Note : The figuresin parentheses the corresponding ratio to the total represent percentage number of patterns (512 ).
Table 9.6 Resultsof weight- degradationtests using Tower of Hanoi position # 3 patterns Best Moves 35 (0.07) 35 (0.07) 0 (0.0) 0 (0.0) 5 (0.01) 0 (0.0) 0 (0.0) 0 (0.0) Long Moves 270 (0.53) 272 (0.53) 0 (0.0) 116 (0.23) 0 (0.0) 192 (0.38) 0 (0.0) 158 (0.30) Regressive Moves 44 (0.09) 44 (0.09) 0 (0.0) 165 (0.32) 0 (0.0) 56 (0.11) 0 (0.0) 52 (0.10) Other Moves 163 (0.31) 161 (0.31) 512 (1.0) 231 (0.45) 507 (0.99) 264 (0.51) 512 (1.0) 302 (0.60)

Degradation to Weight-Matrixs Top half Bottom half Left half Right half Top- left quarter Top- right quarter Bottom-left quarter Bottom-right quarter

: The Aguresin parentheses Note representthe correspondingpercentageratio to the total number of patterns (512).

using a nondegraded network, and then with the different degraded networks . Tables 9.5 and 9.6 show the results of these tests using the 512 pattern variations for position # 3. The results obtained from the eight degradation tests indicate that degradation to different areasof the output . Moreover, the network weight matrix affects the network in different degrees suffers little or no performanceloss when damageis done to the top (W [1- 13, 1- 16]) or bottom ( W [13- 27, 1- 16]) halves of its weight matrix, but suffersa large performanceloss in all other cases . For the tests with pattern # 3, the network seemsto suffer nearly the same loss in performance whether the damageis done to the left half or to the bottom -left or top -left quartersof the output weight matrix. This seemsto indicate that most of the

223

Cardoso& : Parks : ExecutiveFundioning

information necessary to do the third move of the TOH resides in the connections between the output neurons and hidden neurons 1 to 8.

SUMMARY
Under ideal learning and testing conditions, the neural network solved the TOH in seven moves. neural network normal ~ Duzzle ~ ~~ Our ~ ~~ ~~ ~ ~ underestimated ~ -~ ~ ~ ~ ~ - -~- - ~ - -- - ------;

' I controls performance by approximately three moves . This was not unexpected , since a neural network can be very efficient in solving tasks when degradation is not present . Our PlY was used to give us an approximate measure of performance in solving the TOH , both for humans and for the neural net . The total number of errors was inexorably tied to the degradation of the neural network . This was not just a superficial reproduction of the input / output of the data. While it is true that certain aspects of the mathematical algorithms and computer architecture set rigid boundaries of experimentation , there are wide amounts of variability in potential outcomes . The dynamic qualities included varying " " degrees of flexibility in solving the puzzle . This was sometimes represented in the apparent perseveration , as reflected in regressive TOH moves in the neural network . More specifically , the PlY index is a quantitative measure " " that also indicates that the ratio between the best moves and the total number of moves was larger for the normal controls . In addition , we were very surprised that with near- catastrophic degradation of the computer neural network , that is, running the computer program with only a small number of connections intact , the left frontal - damaged group could still solve the TOH in approximately fourteen moves , albeit with more error violations . These experiments were our first attempts aimed at investigating the effects of degradation (radical changes) of the connections of feedforward multilayer neural nets. In our main test , the modified weights were those of the connections between the output and hidden layers . The simplicity of our " statement above - degradation of different areas of the output weight " matrix affects the network in different degrees - might be interpreted as misleading . What did the network learn? What was it trained to do? A simple backpropagation network , like the one used in this work , can learn how to generate a particular output in response to a given input . Perhaps it might be argued that presenting the network with a pattern representing one of the TOH positions , and training it to produce one of the next possible puzzle positions , is not the same as teaching the net how to solve the TOH puzzle . " " The sequence or succession of moves is introduced in the process when the researcher intentionally takes one output pattern generated by the net and feeds it back to its input to obtain yet another response. Nevertheless , it is reasonable to say that the net has learned to generate correct responses, and may be capable of making some generalizations based on what it has learned.

224

NeuropsychologicalTests and Clinical Syndromes

1 0.8

I 0

0.6 0.4 0.2 0

10 11 12 13 14 15

hidden neurons
. neurons -laYei in the hidden Figure 9.9 Plot of the outputactivity generated by the pattern of 3. representative position

Careful investigation shows why the performance of the network was seriously affectedwhen mutilating the left side of the output weight matrix: it was mostly becausethe affectedconnectionswere coming from the active neurons in the hidden layer. Further analysis of the responseof the hiddenlayer neuronsto the pattern of position # 3 (pp3) (seefigure 9.3), shows that the neurons that becomemost active are neurons 3, 5, 6, 7, 8, and 14 (figure 9.9). Becausethe majority of the active neurons (neurons 3,5,6,7,8) are located on the left side of the matrix, changing the weights of these connections would have the largest effect on the performanceof the neural network . Therefore , a combination of factors is responsiblefor the degradation ' of the network s performance . These observations seem to support the view that, at least in our case , forward multilayer neural net will shapea network in which training a feed " " the acquired knowledge is stored both locally and in a distributed manner . In the case of pp3, which can also be interpreted as an individual pattern, thing, or event, only six of the fifteen hidden neuronsrespondto the pattern. This meansthat the net usesboth a local (not every neuron will fire), and a distributed internal (more than one neuron will fire) representation of its " world ." ' Top-down interpretation of the network s learning process will confirm that degradation of connections , by itself, will not necessarilyaffect ' the network s performance . Loss of performance for a given task is most " . These noticeable when degradation is done to the relevant" connections connectionscome from the active areasof the network. Neural networks are capableunder most conditions of creating internal representationsof their " outside world " that enablethem to ably well in the tasksfor perform reason which they are trained. Further evidence of the TOH relationship , using the related Tower of London test, to the left frontal region is found in a brain imaging (single ) study by Andreasenet al. photon emissioncomputed tomography, SPECT (1992). It was shown that normal controls activated the left frontal lobe, -naive and non-naive but two groups of schizophrenicpatients (neuroleptic

225

& : Parks : Executive Cardoso Functioning

. patients with a three-week medication washout) did not activate this area Andreasenet al. pointed out, "The task involves planning (often considered a dorsolateral frontal function), attention (often considereda mesial frontal and cingulate function), and manipulation of shapes(a right parietal function) . .. These results emphasize the importance of conceptualizing cognitive " challengesin terms of the multiple circuits that are likely to be activated (p. 955). Another SPECTstudy with the Tower of London test found activation , et al., 1993). primarily in the left frontal lobe (Morris , Ahmed, Syed This executive function of the ability to plan or anticipate was modeled well in the components of the neural network architecture and during learning trials. In other words, one could say that the input vector is the " where the " TOH ring is and the target vector is the " where I want to move the ring ." The simultaneousintegration of the input and target vectors represent a -type anticipatory or executive function. human Our network also contains attentional elements . In a well- developed neural network model of Stroop test performance , Cohen, Dunbar, and McClelland (1990) viewed attention as a modulation of processing in which additional sourcesof information provide a sustainedcontext for the processingof signals within a particular pathway. In our present model the attentional parametersare served by modification of the delta rule with a " momentum " factor. The gradient-descentalgorithms are subject to incompletion of a task becauseof lowered inertia as a function of the learning trials. By analogy, Caudill and Butler (1990) have indicated that " a sled can overcome small bumps or even short rises in its path if it has generated enough physical momentum to carry it over suchperturbations and to allow it to continue its original direction . . ." (p. 190). The modulation ability to sustain the neural network appears more prominent in the early learning trials. With major degradation of our neural network, there was a negligible in the overall ability to keep focusedon solving the TOH . With our decrease current limitations it is not easy to separateexecutive functions from the attentional functions with the backpropagationrule. Indeed , one of the major tasks for neural network modelers will be to work with the limitations of , many of which are multifactorial in cognitive targeted neuropsychologicaltests . These limitations do not negate the utility of the TO H ' s processing , Butters, ability to delineate differencesin clinical populations. For example Wolfe, Martone, et al. (1985) used TOH proceduresas an index of proce' duralleaming and demonstratedthat early Huntington s disease patients and controls were able to acquire the solution of the puzzle , but advanced ' Huntington s diseasepatients and amnesicsshowed little improvement after concluded that the TOH was of limited repeated trials. These researchers value for investigation of procedural learning becauseof dependenceon problem-solving abilities. Shall ice (1982) basedsomeof his findings on an earlier computer model of the TOH . That artificial intelligence (AI ) model had its origins in a hypothetical serial processing paradigm ( Newell & : Simon , 1972; Simon , 1975).

226

NeuropsychologicalTests and Clinical Syndromes

' This AI computer model, by today s standards , , was limited by its slow speed solve . It could not and small architecture , storagecapacity adequately simple . The reason for most of these limitations was multiple tasks simultaneously ' the program s use of traditional procedural programming languages and methods , where one step follows another in a sequentialrecipe style. In this of , in the program code , type program a minor perturbation in the hardware or in its data, usually causesthe entire program to fail or to stop. In our es its tasks in current model, a connectionist model, the computer process a distributed way. Connectionist architectures can process multiple tasks ably well even after suffering significant simultaneouslyand function reason , Simon made a significant damage to the network, or to its data. Nevertheless contribution to our understanding of executive functioning. In a rigorous task analysisof how to solve the TOH , Simon suggestedfour main es: (1) rote strategy, (2) perceptual strategies , (3) move-pattern approach . last The 4 recursion , the goal-recursion , and ( ) goal strategies strategies strategy, was determined to be the most efficient method of solving the it used " goal and subgoal structures to determine what to puzzle because " do next (p. 273). J. R. Anderson (1993) and colleagues (J. R. Anderson, Kushrnerick , 1993) have provided empirical evidenceto show that , & Lebiere the time to makea move in the TOH is strongly correlatedwith the number of subgoals that must be set before that move. Maximum subgoaling behavior was a characteristicof our neural network of human TOH behavior, whereas the left brain- damaged computer program group had more difficulty in modifying behavior and made more regressivemoves. Acharacteristic of executive functioning is that frontally damaged patients have , and difficulty with cognitive flexibility , considering response alternatives Miotto Morris 1993 & Grattan , et , , , ; Feigenbaum linger modifying plans (Es holds true for al. 1997). This finding of impaired executive subgoaling ' s disease ' , and Huntington , Parkinsons disease patients with schizophrenia , Andrews, Smith, et al., 1996). Moreover, when the problem solving (Hanes , performance complexity of the TOH is increasedby adding another disk . & Fisher 1997 declineswith normal aging (Brennan , , Welsh, ) From a developmental perspective , using a modified Tower of London -injured children and adolescents , Kufera , , Levin, Fletcher procedure in head et al. (1996) were able to demonstratethat this task involves severalcomponents . For example , the number of rules that were broken and the quantity of " " " " was problems solved loaded on a mental construct of planning. Schema " " related to number of problems solved during the first trial and inhibition was associatedwith initial planning time. Some difficulties remain in comparing the TOH and Tower of London tasks owing to validity and standardization issues(P. Anderson, Anderson, & Lajoie , 1996; Kafer & Hunter, 1997). In addition, some evidence suggests only moderate correlations between the two tests ( Welsh , Cartmell, et al., 1997). , Petterson for the visuospatial have does current The parametersto account program " " functions of the TOH task. One of the usesof the context inputs depicted

227

: ExecutiveFunctioning : Parks Cardoso&

in figure 9,1 was to help the neural net during training to distinguish between the different types of legal moves, The network for both the normals and the left &ontal group would have made numerousvisuospatial errors in stacking the rings without this function (illegal state # 5 in figure 9,3 of stackinga larger ring on a smallerring), One type of error that the left &ontal lobe- damagedgroup made in our neural network was a regressive move (#6 in figure 9,3), This could also be called a perseverativeerror, This is the same as the inefficient " perceptual strategy" &om Simon (1974) in which individuals sometimesfind themselvesrepeating the sametask, In spite of the limitations of our current ability to program complex neuropsycholog test performance , we were able to show some of the more salient elementsof executive functioning in the network and approximations in a clinical population, In our laboratories we are programming featuresto articulate more fully the role of working memory and inhibition , A recent theoretical model has implicated working memory in an interactive &ame work of executive functioning and the &ontallobes (Levine & : Prueitt, 1989; Roberts & : Pennington , 1996), In agreement with Cohen and Servan Schreiber (1992), we have designed a rudimentary system using statistical and clinical data to capture those featureswhich are most relevant at a cognitive level analysis" without having to reproducethe entire brain to do so" (p, 117), Future testing of neural network programs that are interfaced by clinical neuropsychological data and in vivo brain imaging, for example , emission PET SPECT functional resonance , ), positron tomography ( magnetic would help to establish imaging, and high-resolution electroencephalography a more solid understanding of the learning and internal representation es and the mechanismsinvolved when dealing with feed forward process nets, For recent PET studies with the Tower of London , multilayer example task found distributed networks of cortical areasincorporating pre&ontal and : Petrides ,& , 1996) in addition to cingulate , premotor parietal areas(Owen, Evans , parietal , and occipital cortices (Baker , Rogers , Owen, et al., 1996), Bakerand co-workers concludedthat a supervisory attentional systemmodel " provided a firm theoretical perspective &om which to regard complex " problem solving by man and machine (p, 524), Currently, brain imaging has ' far exceeded current modelers expectations of demonstrating distributed Baker et al" 1996; Crammond ( , 1997; Frack neuropsychological systems owiak, 1994; Gevins, Leong Smith et al" 1995 Horwitz et , , ; Grady, Maisog, , al" 1994; Grady, Mclntosh, Horwitz , et al" 1995; Parks , Loewenstein , Dodrill , et al" 1988); and executive functions in neuropathological populations : Prueitt, 1993; Weinberger , Parks ,& (Levine , 1993),

ACKNO IWLEDGMENTS

Neural Network Model in

ControlsandLeft FrontalLobePatients Healthy

228

NeuropsychologicalTests and Clinical Syndromes

, by R. W . Parks and J. Cardoso (1997), InternationalJournalof Neuroscience a with . The Tower test NeuroCommunication Research , , used permission variation of the TOH for children, may be found in a subtestof standardized : Kemp, 1997). Comthe NEPSYneuropsychologicalbattery ( Korkman , Kirk , & puterized versions of the TOH , Tower of Toronto , and Tower of London tests may be obtained &om Colorado AssessmentTests , 102 E. Jefferson , Colorado Springs , Inc, , CO 80907, and Psychological Software Services 6555 Carrollton Avenue, Indianapolis , IN 46220.

REFERENCES
. Amm ' canPsychologist Anderson , 35- 44. , 48 , J. R. (1993 ). Problem solvingandlearning . In J. R . : Lebiere . , Kushmerick Anderson . N. & , C. (1993 , J. R ). Towerof Hanoiandgoalstructures . 121 142 . Hilisdale : Erlbaurn the mind . Anderson .), Rules , ( ) NJ (Ed of pp : Validationand : Lajoie Anderson , G. (1996 , V., & ). The Tower of Londontest , P., Anderson . TheClini C411 for pediatric standardization , 10 , 55- 65. Neuropsychologist populations CardSortingtest : Tranel Anderson , D. (1991 , R. D., & ). Wisconsin , S. W., Damasio , H., Jones C411 and & . Journal of frontal lobe damage as a measure of Clini ptrimtntal Neuropsychology performance . - 922 , 13 , 909 : . , Swayze , G., & , P., Cohen , V. W., Flaum , M., Kirchner ,K . Alliger ,R Andreasen , N. C., Rezai -naivepatients with chronic in neuroleptic andin patients . (1992 O'leary, D. S ). Hypofrontality -photonemission and with xenon133single : Assessment tomography computed schizophrenia - 958 . . Archives theTowerof London , 49 , 943 of General Psychiatry ., Yetkin : Lobeck . M., Bernard in, l., Grafman Arnett . L. (1994 . J , F. Z., & ). Relationship , P. A , Rao ,S with in patients andWisconsin CardSortingtestperformance frontallobelesions between 44 420 425 . . sclerosis , , Neurology multiple . S. J., & . J., Frackowiak : Robbins . D., Owen , Baker .R , C. D., Dolan ,R ,R , A. M. Frith , S. C., Rogers . : a PETstudyof the Towerof Londontask T. W. (1996 ). Neuralsystems by planning engaged - 526 . , 34 , 515 Neuropsychologia . Neuropsychologia in frontallobedisease effects behavioral , 6, 53Benton , A. (1968 ). Differential 60. : An functionskills : Fisher Brennan , C. B. (1997 ). Aging and executive , M, Welsh , M. C., & . Perceptual and Motor Skills of a community - dwellingolder adult population , 84 , examination - 1197 . 1187 . (1985 : Cermak Butters . L. S , E., & ). Memorydisorders , M., Granholm , N., Wolfe , J., Martone 's disease . andprocedural : Verbalrecall with Huntington associated , verbalrecognition memory . 23 729 743 , , Neuropsychologia . AI Expert theTowers Cardoso , October , 49- 53. , J. (1991 ). Revisiting of Hanoi . . Cambridge : Butler Caudill , MA: MIT Press , M., & , C. (1990 ). Naturally systems intelligent C41tions . : Theory : Rurnelhart , architectures , andappli Chauvin , Y., & , D. E. (1995 ). Backpropagation . Hillsdale : Erlbaurn , NJ :A es : McClelland Cohen , J. L (1990 , J. D., Dunbar , K., & ). On the controlof automatic process - 361 . Review . Psycholog of the stroopeffect account distributed , 97 , 332 ;C4II processing parallel -Schreiber in psyChia to neural networkmodels : Servan Cohen , D. (1992 , J. D., & ). Introduction 22 113 118 . . Annals , , try Psychiatric

229

: ExecutiveFunctioning : Parks Cardoso&

Crammond . D. J. (1997 : Neverin your wildestdream ) . Motor imagery . Trends in Neurosciences , 20 , 54- 57. Crockett , D. J., Bilsker : Kozak , D., Hurwitz , T., & . J. (1986 ). Clinicalutility of threemeasures of frontal lobe dysfunction in neuropsychiatric . Intmranonal ] OUmRI , 30 , samples of Neuroscience - 248 241 . Dodril L C. B. (1997 . TheClinical ). Myths of neuropsychology . , 11 , 1- 17 Neuropsychologist -striatalsubstrates Es : Grattan , P. A , & , L. M. (1993 ). Frontallobeandfrontal linger for different - 28. formsof human . Neuropsychologia , 31 , 17 cognitive Rexibility Fletcher , J. M. (1996 functions in children : Introduction ). Executive to speaalseries . Dtveiop mmtRI , 12 , 1- 3. Neuropsychology . S . J. (1994 Frackowiak , R ). Functional . Trends in mappingof verbalmemoryand language Neurosciences . - 115 , 17 , 109 Gevins , A , Leong : Du , M. E., Le , J., & , R. (1995 ). Mappingcognitivebrainfunction , H., Smith -resolution - 436 with modern . Trends . high , 18 , 429 electroencephalography in Neurosciences Glosser : Goodglass , G., & in executive , H. (1990 controlamongaphasic ). Disorders and other - 501 braindamaged . ] oumRl tmdErperimmt R I Neuropsychology . of Clinical , 12 patients , 485 , C. L., Maisog , B., Ungerleider Grady , L. G., Mentis , M. J., Salerno , J. M., Horwitz , J. A , Pietrini , -relatedchanges P., Wagner : Haxby , E., & , J. V. (1994 in corticalblood Rowactivation ). Age of faces andlocation . The - 1462 . duringvisualprocessing , 14 ] oumRl , 1450 of Neuroscience . , Horwitz , C. L , Mcintosh , A. R , B., Maisog Grady , L G., Mentis , M. J., , J. M., Ungerleider Pietrini : Haxby , P., Schapiro , M. B., & , J. V. (1995 in human reductions ). Age-related recognition dueto impaired . Science - 221 . , 269 , 218 memory encoding Hanes .R . , Andrews ,K : Pantelis , D. G., Smith , D. J., & of executive , C. (1996 ). A briefassessment controldysfunction : Discriminant of planning validity andhomogeneity , and Ruency , set shift - 191 measures . Archives . , 11 , 185 of Clinical Neuropsychology Hoist : Vilkki , P., & , J. (1988 on perfonnance ). Effectof frontomediallesions on the Strooptest andword Ruency tasks . ] oumRl tmdErperimmt R I Neuropsychology . of Clinical , 10 , 79 Kafer . L, & -solvingtasks : Hunter ,K , M. (1997 ). On testingthefacevalidity of planning / problem in a normal . ] oUmRl Rti I Neuropsychological On R - 119 . of theIntern , 3, 108 population Society Korkman : Kemp . (1997 , M., Kirk . U., & mRnURI . SanAntonio ,S ). NEPSY , TX: Psychological . Corporation Levin , H. S., Fletmer , J. M., Kufera , J. A., Harward , H., UUy , M. A , Mendelsohn : , D., Bruce , D., & of cognitionmeasured ). Dimensions the Tower of London andother Eisenberg , H. M. (1996 by -injureddti1dren tasks in head andadolescents . Dtvelopment R I Neuropsychology cognitive , 12 , 17 34. Levine , D. 5., Parks : Prueitt , R. W., & , P. S. (1993 andtheoretical issues in neural ). Methodological networkmodels of frontalcognitivefunctions . Inttrnil Honal , 72 , 209 Journal of Neuroscience 233 . Levine : Prueitt . (1989 , D. 5., & , P. S effects of frontallobedamage ). Modelingsome : Noveltyand - 116 . Neural Networks . , 2, 103 perseveration I assessment Lezak . M. D. (1995 Cil .). NewYork ), Neurops : Oxford University (3rd ed Press . ,Vchologi . P. (1987 Hon ,R to compu with neural ) . An introduc nets . IEEE ASSP Lippmann Hng , April Magazine 4- 22. Reprinted in Artificial NeuralNetworks : Thtore H Cil I Concepts . WashingtonD.C.: IEEE , , 36:54 SodetyPress Computer

230

Tests andClinical Neuropsychological Syndromes

. (1993 . G., Ahmed : Toone of planning Morris , B. K ,R , S., Syed , G. M. S., 6 ). Neuralcorrelates the Tower of London test . 31 , , 1367 R ability: Frontallobe activation Neuropsychologi during . 1378 . (1997 . G., Miotto, E. C., Feigenbaum : Polkey Morris , C. E , J. D., Bullock , P., 6 ,R ). The effect -lobe lesions -subgoal in humans . con Aid on planning of goal ability afterfrontaland temporal - 1157 . , 35 , 1147 R Neuropsychologi . wounds . London : Oxford University Press Newcombe , F. (1969 ). Missile of thebrain -Hall. . Englewood : Prentice Newell Cliffs : Simon , NJ , A., 6 , H. A (1972 ). Human problem solving and : Robbins Owen , J. J., Sahakian , B. J., Polkey , C. E., 6 , T. W. (1990 , A M., Downes ). Planning in man . 28 1021 1034 . frontal lobe lesions , , R memory Neuropsychologi working following spatial for a two - stagemodelof spatial : Petrides Owen , M. (1996 , A M., Evans , A C., 6 ). Evidence . : A positronemission within the lateralfrontal cortex study memoryprocessing tomography Cerebral Cortex , 6, 31- 38. -Wesley andneural networks . New York : Addison . Pao , Y-H. (1989 ). Adaptive recognition pattern : distributed and executive Parks . W., 6 : Cardoso , J. (1997 ,R ) . Parallel functioning processing . International Towerof Hanoineuralnetworkmodelin healthycontrolsand left frontallobepatients - 240 . , 89 , 217 Journal of Neuroscience . , Barker Parks . , Loewenstein , W., Yoshii , F., Chang , A., Apicella , A., , D., Dodrill, K ,R , J., Emran : A PET of a verbalfluencytest . (1988 metabolic effects : Duara Sheramata , W., 6 ,R ). Cerebral - 575 . andExperimental . Journal scan , 10 , 565 of Clinical study Neuropsychology in verbal of habitual Perret , E. (1974 ). Theleft frontallobeof manandthe suppression responses 12 . . 323 330 behavior , , Q Neuropsychologi categorical framework for examining Roberts : Pennington , B. F. (1996 , R. J., 6 ). An interactive prefrontal - 126 . es . Developmental , 12 , 105 Neuropsychology cognitive process . A. W., 6 . K., Lehman : Stilson , D. W. (1980 ,R ,R ). The utility of the Robinson , A. L., Heaton . Journal andlocalizing frontallobelesions Wisconsin CardSortingtestin detecting of Consulting . - 614 andClinical , 48 , 605 Psychology in planning . Philosophical Transadions Shallice , T. (1982 ). Specific of theRoyal Society impairments . London 298 209 199 , , of . Cognitive of problem Simon , , H. A (1975 ). The functional Psychology solvingskills equivalence - 288 7 . , 268 . New York: Church ill Uvingston . : A clinical Walsh . W. (1978 ,K ). Neuropsychology approach . Journal to the prefrontal cortex . (1993 , D. R ). A connectionist of Neuro Weinberger approach . , 5, 241 253 psychiatry esaretapped : Stein Welsh , T., 6 , M. (1997 , M., Petterson , N., Cartmell ). Whatcognitiveprocess ? the International of Hanoi and London the Towers , 3, 29. Journal Society of Neuropsychological by

: Executive f ' mdioning Cardoso & : Parks

Neuronal Prefrontal

Network Models De6 ~its

of Acalculia

and

Stanislas Dehaene , Laurent Cohen , and Jean- Pierre

Changeux

How do we recognize that 20 + 21 = 91 is an " obviously" false operation? What are the cerebral circuits involved in mathematical reasoning ? What of cell firing encodethe complex stepsof calculationsinvolved in sequences , say, 13 x 207 And how do localized brain lesions disrupt these computing ? This chapter focuseson neuropsychologicalstudiesand neuronal operations network models of some of the most evolved functions of the human brain: arithmetic, mathematics , reasoning , and abstract rule formation. The neurobiological databasethat would support accuratemodeling of these functions at the neuronallevel is still lacking. Nevertheless , there is a growing body of evidence on the large scale networks underlying arithmetic, both from single-casestudiesof brain-lesionedpatients and from brain imaging studies in normals . The development of adequatemodels of the large networks of brain areasinvolved in calculation and reasoning , together with more speculative level at the neuronal , may ultimately set the stage for finer modeling . neurobiological experiments NEURONAL MODELS OF COGNITIVE FUNCTIONS Before building neuronal models of cognitive functions as complex as the , the theoretical basesof such an enterprise must production of mathematics be set. With its billions of cells and connections , woven by evolution over the brain is the most hundredsof millions of years , complex organ we know of. Relating cognitive functions to their neuronal substratesis fraught with epistemologicaldifficulties. In an article that appearedin 1989 in the journal , two of us presentedwhat we believe are the minimal preconditions Cognition for neuronal modeling of cognitive functions (Changeux & : Dehaene , 1989). Two conditions are crucial: the distinction of levels of cerebralorganization , and the modeling of modesof transition between theselevels. The first theoretical enterprise , which is essentially qualitative, consistsin a into hierarchicallevels of organization, the architecture cleaving complex in of which coincided the course of evolution, with the emergence , appearance of new functions. The first level, that of molecules , includes genes and proteins. Intuition notwithstanding, the molecular level is often directly

relevant to some cognitive functions, as attested , for instance , by the key role of pharmacologicalagents in psychiatric diseases . Immediately above the molecular level stands the cellular level, which is characterizedby an anatomical disposition unique in the organism : the ability of nerve cells to form topological networks defined by their multiple axonal and dendritic projections. At this level, the elementaryfunctions of some cells may have a 's measurable influenceon an organism behavior. This is especiallytrue in the caseof identifiable neurons such as the Mauthner cell, which governs the fleeing behavior of fish. The level of organization immediately above the cellular level is that of small ensemblesof cells , or the circuit level. Each circuit is composed of several hundred of thousands of neurons interconnected for performing a defined function. One of the simplest examplesis the network of motor cells in the spinal cord of a primitive vertebrate , the lamprey. The circuits that ' govern the lamprey s different swimming modes are known in great detail, to the point that it is possible to simulate on computer the oscillatory behavior of identified neurons and to predict the effect of lesions or of pharmacologicalagentson behavior. At a higher level, the circuits connect in complex assemblies whose functions , in the human brain are only beginning to be delineated . It is well known, for instance , that visual functions are representedin a posterior network which may be subdivided in further assemblies coding mainly for the position and identity of visual objects. At a yet more integrated level, a set of pre&ontal cortical areasis crucial to the planning of actions in time and , and contributes to the construction of motor programs and of novel space . The surfaceof the pre&ontal cortex sharply increasedin evolution hypotheses , reaching its maximum in man where it occupies almost a third of all cortex. It therefore seemslegitimate to propose that the pre&ontal cortex " " the brain of , which characterizes participates in the architectureof reason Homosapiens . Here, the cleavageinto levels of organization clearly remains approximate and will have to be progressively refined as new techniques allow for more preciseanalysesof cortical activity in humans . Neuronal modeling of any cognitive function therefore starts with the choice of the levels of organization relevant to this function. In the caseof number processingand pre&ontal functions, we will not addressthe molecular or cellular levels. However, the models that we examine do lead to specific predictions concerning the cellular and molecular learning mechanisms that underlie cognitive development . The levels that we consider here concern " the neuronal circuits and the circuits of circuits" (assemblies ). We will see that it is possible to outline a plausible diagram of the main areas for number processingin the human brain, to set out a neuronal model for the development of a crucial component of this diagram (the quantity representation ), and finally, to speculateon the role of pre&ontal areasin more advancedcalculationand reasoningabilities.

234

a J Tests andClinical Neuropsychologic Syndromes

The secondprecondition for building neuronal models of cognitive functions consists in defining the transitions from one level of organization to the next. In the context of the application of evolutionary theory to the : Danchin, 1973; , Courrege ,& developmentof the nervous system (Changeux 1983 & : Danchin 1976 , ; Heldman : Patte ; Changeux , ,& , Changeux Changeux 1984; Changeux & : Dehaene , 1989), we have proposed an acccount of the transitions between levels with a variation-selection scheme , also called a , 1978, 1987). At eachlevel, generalizedDarwinian scheme(seealso Edelman a " generator of diversity " introduces variations in functional organization which are selectedand propagated or eliminated depending on their relevance to higher levels. At the circuit level and beyond, this internal evolution is epigenetic and does not require genomic modifications . In the course of development , the variable elementsare the neurons themselvesand the topology of their connections . The selective stabilization of synapses , together with the elimination of connections es a circuit adapted to a given function. At , establish higher cognitive levels, the extension of the evolutionary paradigm suggests that the brain functions on a projective mode by permanently producing " tentative representationsor " prerepresentations variable in time and space that are adopted or rejected by the organism based on their adequation to the external world . In this " mental Darwinism" framework , development and can be described as a reduction of a vast initial repertoire learning progressive of prerepresentations . We shall seebelow severalexamplesof how this conception of development applies to the acquisition of elementary numerical . knowledge and to simple reasoning A NETWORK OF AREAS FOR NUMBER PROCESSING , the first step in achieving a neuronal According to the above principles model of number processingand calculation consistsin identifying the gross network of cortical areasinvolved and their individual functional contributions . Drawing from studies of mental chronometry, neuropsychology , and brain imaging, we have recently outlined a tentative model of the cerebral : Cohen 1995). , 1992; Dehaene& circuitry for number processing(Dehaene Our " triple-code model" acknowledgesthat there are three main ways in which numbers can be manipulated mentally: (1) a visual arabic code , (2) a verbal code , and (3) an analogicalmagnitude code. In the visual arabic code , numbersare encodedas strings of digits on an internal visuospatial scratch pad which, by analogy with word processing , is called the visual number form. In the verbal code of number words. , they are encoded as sequences , numericalquantities are representedby local Finally, in the magnitude code distributions of activation on an oriented number line. Numerical relations ; such as the fact that 9 is larger that 5, are then implicitly representedby proximity relations on the number line.

235

Dehaene : Changeux : Acalculia and Prefrontal ~ , Cohen , &

cit.4

The model assumesthat the three cardinal representationsare linked by direct transcoding routes that enablenumbers to be rapidly re-encodedin a different format. For instance , a printed arabic numeral is initially identified and encodedin the visual arabic numeral form. If the task requires assessing its magnitude , then its identity will be transmitted and re-encoded on the . If, on the contrary, the task requires reading the quantity representation numeral aloud, then its identity will be transmitted to the verbal system where an appropriate sequenceof words will be composed . In the course of more complex tasks , such as a multidigit multiplication, several such transcodings may take place back and forth between the three cardinal . representations A crucial hypothesis of the model is that each of the internal numerical is well suited for a different set of functions. The visual number representations form, holding a representation of the spatial layout of digits, is well suited for performing spatially organized mental calculationswith multidigit arabic numerals . The verbal format, on the other hand, is well suited for a stored table of rote arithmetic facts , which we suppose to be accessing encoded in the form of short sentencesin verbal memory (e.g ., " two times three, six" ). Finally, the analogicalquantity representationis ideal for representing , such as knowing approximate relations between numericalmagnitudes that 11 is about 10, or that 7 is slightly larger than 5. Thanks to several single-case and brain imaging studies , reviewed in Dehaeneand Cohen (1995), a schematicmodel of the anatomicalnetwork of areasunderlying the triple -code model may be proposed (figure 10.1). The visual number form correspondsto the endpoint of a cascade of areasculminating in the medial occipitotemporal region of both hemispheres . The analogical is quantity representation computed by areasin the vicinity of the . Finally, the verbal parieto- occipitotemporal junction of both hemispheres word frame is computed by the perisylvian languageareasof the left hemisphere . We postulate that homologous left and right hemisphereregions-

of the anatomical and functional drcuits postulated Figure 10.1 Schematic diagram by the -codemodel to underlie number & Cohen . 1995 ( Dehaene ). triple simple processing

236

Tests andClinical Syndromes Neuropsychological

magnitude representation

- are the two visual number forms and the two quantity representations connectedinto a single functional unit by transcallosalfibers. Within the left , we also supposethe visual, verbal, and quantity representations hemisphere to be fully interconnectedby dedicatedneural transcoding pathways. Support in favor of this network organization of numericalrepresentations has come from the observation of highly selective deficits of number processing in casesof focal brain lesions . For the sake of brevity , only a few such observations are discussedhere (see Dehaene 81 : Cohen, 1995). First, studies of split-brain patients in whom the corpus callosum has been cut : Hillyard , 1971, Gazzaniga81 : Smylie, 1984; Seymour (Gazzaniga81 , Reuter Lorenz : Dehaene , and Gazzaniga , 1994) or lesioned(Cohen 81 , 1996) conform in great detail to the predictions expected from the model. When digits are ' flashed in their right hemifield , therefore contacting the patients left hemisphere visual system , split-brain patients can name them, judge their numerical size , and calculatewith them. This is consistentwith the presenceof a full network comprising visual, verbal, and quantity representationswithin the left hemisphere . The situation is different when digits are flashedwithin the left hemifield and therefore contact the right hemisphere . Split-brain patients still whether two may judge digits are the same or different, point to the appropriatedigits in an array, and And the larger of the two digits . This conArms that the right hemispherecan both identify digits and associatethem with the corresponding numerical quantity. However, the patients cannot read the very samestimuli aloud, nor can they calculatewith them, a deficit which is predicted in our model by the lack of a verbal representation of numbers in the right hemisphere . A similar dissociation between preserved of numerical comprehension quantities and impaired reading and calculation has also been observed in several patients with extensive left hemisphere lesions (e.g . Cohen :, Verstichel, 1994; Dehaene81 : Cohen, 1991; , Dehaene , 81 Grafman et al. 1989 . , Kampen , Rosenberg , ) , A second type of patient that yielded valuable data suffered from a left medial occipitotemporallesion, resulting in a severedeficit of reading called : Dehaene , 1995; Dejerine, 1891, 1892). In our model, pure alexia (Cohen 81 such a lesion is predicted to disrupt the formation of the left visual number form, while sparing completely the right visual number form. As can be seen in figure 10.1, this should have the effect of disconnectingthe verbal system from direct visual inputs, while leaving the quantity representationappropriately connectedto visual inputs via the right hemispherevisual system . The observed impairmentsconform to this pattern. Patients with pure alexia are largely unable to read arabic digits aloud, but they can make very accurate larger/ smallercomparisonswith the samestimuli, indicating a well-preserved . Likewise , they are unable to perform even comprehension of quantities calculations with written arabic stimuli (e.g ., 2 + 2), becausetheir left simple verbal cannot be informed of the identity of the system directly hemisphere visual operands . Yet they perform the very same calculations with perfect accuracy when the problems are presented orally, thus proving that the

237

Dehaene : Changeux : Acalculia andPrefrontal DeAcits , Cohen ,&

verbal representationand calculation modules themselvesare intact and can still be accessed from the auditory modality . Overall, these patients provide a striking illustration of the speciBc deficits that can occur when a single circuit or connection is disrupted within the large-scalenetwork of areasfor number processing .

A NEURONAL MODELFORTHEDETECTION AND INTERNAL AnON OF NUMERICAL REPRFSENT Nn Tl~ QUA


The triple -code model accountsfor aspectsof mental arithmetic and acalculia at an abstract level of organization that correspond roughly to that of " circuits of circuits." The internal representationsin the model are currently describedverbally and without explicit simulation. Ultimately , it would be desirableto understandtheir implementation within local neuronal networks. As a step in this direction, we have proposed a detailed neuronal model for a crucial part of the triple-code model, namely, the quantity representation : Changeux , 1993). (Dehaene& In many nonverbal organismssuchas rats and pigeons , and even in human babiesin their first year of life, the quantity representationseemsto be functional in the absenceof any symbolic visual or verbal representation of numbers . Young infants can discriminate small numbers of objects , such as two vs. three, regardlessof the modality of input (e.g . Starkey & : Cooper, : Mehler, 1991; Starkey : Gelman 1980; Bijeljac-Babic ,& , Bertoncini, & , Spekle , 1983, 1990). Severalanimal speciescan determine the numerosity of collections of objects and learn to associate specificbehaviors with specificnumerical For Meck . instance and Church (1983) taught rats to press , quantities one lever when they heard two sounds and another level when they heard , regardlessof their duration. Importantly , animals can use this eight sounds representation to perform larger/ smaller comparisons (e.g., Mitchell , Yao, -Rumbaugh : Hegel, 1987), and in Shennan , et al., 1985; Rumbaugh , Savage ,& seem to be the same kind of so they using doing analogical representation . In number comparisons that adult humans use , both animals and adult humans are affected by a distance effect (comparison speed and accuracy ) drop for increasingnumerical distancebetween the two comparednumbers and by a magnitude effect (for equal distances , speedand accuracydrop for ). Theseparallelshave led to the hypothesis that increasingly larger numbers humansand animalsare endowed from birth with a preverbal representation of numerical quantities which, only in humans , later connects to visual and verbal symbolic codes to form the complete triple -code model (Dehaene , . : Gelman 1992 1992; Gallistel & , ) Given such data, our goal was to provide a plausible neuronal network that would reproduce the functions of the preverbal quantity representation found in animalsand in young infants. Sucha network, when presentedwith a set of visual or auditory objects , should detect approximately how many

238

andClinical Syndromes NeuropsychologicalTests

VI . usllnput

50input clusters

88 . 8 888 . . . ' I 1 . : 18 11 1 I . . : _ 8 : : I ' 4 ' , " , " " .. ' 8 8 8 . . 8 e A A . /\-i ~ 15 ckJst & r s with 15 nume ~ U1res clust ing
Object Location and Normalization

Numero8lty Detection

9 x 50array of DOG filters

Figure 10.2 Detailed connectivity of a neural network model capable of extracting the & : Changeux . 1993). numerosity of a set of objects ( Dehaene

objects are present and encodethat approximate number on a representation . compatiblewith the experimentally observeddistanceand magnitude effects The first crucial component in our model is a numerosity detection system : Changeux (Dehaene& , 1993; see figure 10.2; see also figure 10.4 A ). This network : (1) an input " retina" on prewired comprisesthree distinct modules which visual objects of different sizes can be presented ; (2) an intermediate topographic map of object locations, on which eachobject, independently of its original size , is representedby an approximately constant pool of active neurons , and (3) an array of numerosity detectors that compute the total activation over the location map, thus producing a quantity highly correlated with the number of objects on the retina. This architecture defines a . A sequential parallel algorithm for enumerating objects of different sizes was also defined and connected to auditory input numerosity detectors with matched to the Horn the location weights weights map. As a result, the numerical output of the system was amodal . Thanks to the size normalization operation, all visual and auditory objects were counted as " one." Figure 10.2 shows in greater detail our computer implementation of this neuronal network model. Each neuronal circuit is composed of many units " " called neuronal clusters , each representinga small ensembleof about 100 or 1000 highly interconnectedneurons with similar responsecharacteristics . This reproduces to some extent the columnar organization of the cortex castle , 1978; Goldman-Rakic ( Mount , 1984). dusters are connected by bundles of synapses(seeDehaene& : Changeux , 1993, for a precisedescription of the connectivity used ).

239

Dehaene : Changeux : Acalculia and Prefrontal Delicts , Cohen , &

Average activity 1 1

0 1 2 4 3 5 6 7 8 9 11 12 14 13 15 ,10
cluster

" to different of fifteenoutputunits ("numerosity detectors Figure 10.3 Simulated responses ) numbers of inputobjects 6 : Changeux ( Dehaene , 1993 ).

In one simulation, we presented2500 sets of one to five blobs of different sizeson the input retina. As shown in figure 10.3, eachoutput cluster of the network activates reproducibly when a certain given number of objects is " " . present at input . We therefore termed these units numerosity detectors Similar results were obtained with objects presentedin the auditory modality , or with simultaneousauditory and visual inputs. Eachnumerosity detector , regardlessof their input responds to a fixed total number of objects modality . Figure 10.3 makesit clear that the coding of numbers in this network is quite different from the symbolic and discrete representationsused in computers . Owing to an intrinsic variability in the normalization operation, the activations evoked by different numbersof input objects overlap in part. For instance , some numerosity detectors react when either four or five objects are present , indicating that number coding is approximate and analogical rather than digital . Note that the activation overlap decreases with numerical distance (distance effect). Furthermore , for equal numerical distances , the activation overlap increases with numerosity: the activation peaksevoked by 4 and 5 overlap more than those evoked by 1 and 2 (magnitude effect). In fact, the network codes numerosity with fairly discrete peaks only up to three objects , thus providing a natural account of the so-called subitizing range of one to three items, over which humansare very accurateinassessing numbersof objects without counting (Mandler & Shebo , 1982; Oehaene & Cohen 1994 . The of , ) decreasingprecision numerosity coding in the network derives directly from the mannerin which numerosity is estimatedas a sum of approximately constant activations associatedwith each counted , the higher the variance in the final object: the more terms in this sum estimate .

240

Tests andClinical Syndromes Neuropsychological

CONDm ONI NG TO NUMEROSITY


In order to directly simulate animal conditioning experiments , we have connected the numerosity detectors to a number of output clusters (figure 10.48 ). The numerosity -to - output connections are variable and their efficacy is determined by a Hebb rule modulated by a reward signal (Dehaene & Changeux , 1989 , 1991 , 1993 ). Initially , these connections have a random efficacy and the simulated organism therefore responds randomly . When the response happens to be adequate for the desired behavior , it is selected by giving the network a positive reward which has the effect , thanks to the Hebb rule , of stabilizing recent neuronal activity and therefore of increasing the chances that this correct behavior will be reproduced later on . When the response is inadequate , a negative reinforcement signal is sent which has the effect of destabilizing recent activity and therefore favoring the selection of a different response on later trials . We trained the simulated network on a numerosity discrimination task similar to a task used with rats. For 2 vs. 4 discrimination , for instance, when two objects were presented at input , the correct response was to activate output cluster A , and when four objects were presented , to activate output cluster 8 . The network was trained three times with each pair of numer osities between 1 and 5 ( 1 vs. 2, 1 vs. 3, 2 vs. 3, etc.). In all cases , fewer than es in 300 trials were necessary to reach a stringent criterion of twenty success a row , and asymptotic performance varied between 78.5 and 97.8 percent correct . Two fundamental characteristics of animal numerical conditioning were reproduced by the model . First , the learning time was longer , and the final performance was worse , when the numerical distance between the two discriminated numbers decreased (distance effect ). Thus , it took 285 trials to learn to discriminate 4 vs. 5 with 78.5 percent final accuracy, but only 81 trials to learn to discriminate 1 vs. 4 with 96.8 percent final accuracy . Second, the magnitude effect was also found : it was easier for the network to learn to discriminate , say, 1 vs. 2 than 4 vs. 5. Figure 10.5 shows the connection strengths developed when learning various numerical discrimination tasks. For the discrimination of 2 vs. 4, for instance, the numerosity detectors that were regularly activated by the presentation of two objects developed excitatory connections toward the " " output cluster associated with the response 2. Similar connections were " " found for the response 4. These connections were not strictly limited to the numerosities 2 and 4, but spread to neighboring numerosity detectors , thus providing the network with properties of robustness and generalization . Why are these simulations relevant to the neuropsychology of number processing ? A frequent finding in patients with acalculia is that the ability to approximate numbers can be preserved even in the face of a complete disruption of exact calculation abilities . For example , patient N .A .U ., who could not calculate 2 + 2, nevertheless knew that it could not make 9 (Dehaene & Cohen , 1991 ). He also had excellent knowledge of proximity relations

241

: Acalculia and Prefrontal Deficits Oehaene : Changeux , Cohen ,&

A. Numerosity Detection System

Ic-.nory 1 Itory kfjut 11 - I Auc : )1 Echo Q kJput . .J ~ Ject Location ~ I Vl8wI I ~ & -~ & 1r 8nd 1Io1m8 ~ lIon Q I I~ I

N .....'O8iIy ~ -:w .

C. Comparison of two numeroslt /es

I I

Poi ' :': : : " A ' C ; ~ :J ' 0'

I I

[ ~~

between quantities , such as knowing that 12 is not very different from 8 or 10. Other patients have shown a proximity effect on naming errors , for " " " " " " instance, frequently naming digit 2 as three or four , but never as nine : Aliminosa , 1993 ; Campbell & : Clark, 1988 ). (e.g ., Macaruso , McCloskey , & This evidence suggests that an analogical representation of numerical quantities , such as the one provided by the numerosity detectors in our model , ' underlies the patients judgments . It also supports a postulate of the triple code model that symbolic verbal and digital representations of number are dissociable for core knowledge about approximate numerical quantities . In the animal and human brain , the size normalization and numerosity detection operations postulated in the model may be performed by the dorsal stream of occipital and parietal areas, which are known to participate

242

Tests andClinicalSyndromes Neuropsychological

B. Output andConditioning by Reward -

: : :.

:l-.. -v ' I ' Poi :' : " A " C ; ~ :J ' 0' I

A .. ~ .. ....lion Loop

,
- - -- -- - - -- - -- - -- - -- - - - - - - - -- - -- - ~ 01 ' . - .~ tJ INor ..::' O8lyJ (In ) Figure 10.4 (continued
Plly81c8 ~ ~ ~ ~ 01file 8CIIon

[~ ~

/- I ~ I Act M8tc # tttg I ~ "9I -, ~

[ =:~

.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

243

: Changeux : Acalculia and Prefrontal Deficits Dehaene , Cohen , &

1 12 0 1 2 3 4 5 8 7 8 g 11 10 12 14 13 15 + 4
0 4 + 4 0 4
+ 4 0

,.~ I,' ,'.' ." I .


I ,,' I I , " ' I,' ,~ - .

+ 4 , ~ I ' I ~ , 0 ~ ' " 1 ~ , , ' I I . . 4 + 4 0


, -4 ......
-

\ , . I I \ I ~ J

. -

in the coding of object location and size . In the courseof evolution, the location map may have originally evolved for the parallel computation of object location and size for the purpose of guiding motor gestures . The numerosity detection network may later have preempted part of this spatiomotor . The close links between numerosity extraction and location guiding system coding may explain why bilateral inferior parietal or parieto-occipitotemporal areasactivate during various number-processingtasks (e.g . Roland & : Friberg, 1985 ; Dehaene , Tzourio, Frak , et al., 1996; Dehaene , 1996), and in these areas the left are a critical site for acalculia , , (e.g ., why hemisphere Gerstmann , 1940). At a somewhat homologous location in the associative cortex of the cat, Thompson , Mayers, Robertson, et al. (1970) have recorded neurons that fired after a fixed number of auditory or visual only single stimuli were presented . These neurons thus appeared quite similar to the ' numerosity detectors postulated in our model. The authors findings, however . , have not yet beenpursuedwith modem neurophysiologicaltechniques LEARNING TO ORDER NUMEROSITY DETECTORS The numerosity detection network mimicks an initial stage of human or animal numericalcognition. However, animalsand young children are also able to choose the larger of two sets , an ability that the above network cannot one at any given time. simulatemerely because only numerosity is processed To simulate the development of number comparison abilities, we therefore introduced an additional memory circuit that could maintain, in short-term memory, a representationof the preceding numerosity while a new one was . Two novel neuronal networks were introduced (figure being processed 10.4C). First, a short-term memory network was connected to the numer. It comprised an array of clusters with strong recurrent osity detectors connections , capableof maintaining a sustainedactivity after the excitatory : Changeux stimulus that activated them had disappeared(Dehaene& , 1989, " remanent " in its 1991, 1993). Eachmemory cluster therefore kept activity a trace of the preceding activity level of the corresponding numerosity detector. Second , we also introduced a point -to -point matching network, the clusters of which activated only when the two numerosity and memory clustersthat were topographically connectedto them becameactive. The memory and matching networks, just like the numerosity detectors , were connected to output clusters with connections of variable strength modifiable by the Hebb rule. In order to teach the network a larger/ smaller comparison task, we successivelypresentedon the input retina a set of nl , and then another set of n2 objects objects for thirty -two simulation cycles for sixteen cycles . The simulatedorganism had to activate an output cluster A when nl was larger than n2 , and another output cluster B when nl was smaller than n2 . On each trial, the network received a positive or negative reward depending on the correctnessof its response , and connectionswere

245

: A('~lnJlia andPrefrontal Defidts Dehaene : Changeux , Cohen ,&

modified accordingly. After about 300 trials, the network reachedan asymptotic , the distance and performance of 88.9 percent correct. Once again , we were magnitude effectswere found. Without entering further into details able to simulate a large number of experimental results on sequentialorder learning in animals and children, including the so-called transitive inference tasks (Bryant & : Trabasso : Chalmers , 1971; McGonigle & , 1977; von Fersen , Wynne, Oelius, et al., 1991). The type of learning by reinforcement that we have used in our simulations . However provides a natural accountof animal conditioning experiments , young children do not require any explicit conditioning to acquire . At fourteen months of age , they concepts of larger vs. smaller numbers choose the of two sets & : Adams 1987 ( , ; see spontaneously larger Sophian also Cooper, 1984). As a first step toward simulating suchauto-organization, in the most elaboratedversion of our model we have introduced an architecture in which reward signals are not provided by an external teacher , but are generatedinternally by an autoevaluation loop (figure 10.40 ). The idea is that the simulated organism plays with a set of objects by randomly : adding one object, or removing one object. choosing one of two actions The concomitant variations in numerosity are noticed by the numerosity detection network. By comparing the previously memorized numerosity with the new one, the organism attempts to reconstruct , a posteriori , the action that was performed. An internal comparison module evaluates the . When they match , an similarity of the reconstructed and actual actions internal positive reward is generated by this autoevaluation loop. On this basis , the system progressively discovers that an increase in numerosity , and that a decreasein numerosity implies that an addition was performed . implies a subtraction We have simulated this architecture and shown that indeed , based solely on a correlation between actions and the corresponding changesin numer, first between consecutive osity, the system can discover ordinality relations numerositiesand, by later generalization . , between any pair of numerosities Thus, our network implementsin a working simulation Cooper' s suggestion that " it is the relationship between the numerosity detector states and the effects of addition and subtraction that give rise to the notions of more and " less . From the child' s point of view 'more' is invented in this process . (Cooper, 1984, p. 166). Our simulation can be viewed as an existenceproof that such a learning mechanismis feasible . Whether children and animals use a such mechanism to learn actually ordinality , or whether they have more innate knowledge about the ordering of numerosity detectors , however , remainsto be conclusively studied. A HIERARCHY OF NETWORKS FOR EXACT CALCULA nON

246

andClini~ 1Syndromes NeuropsychologicalTests

of the left hemisnhere circuits of areasimolicated in various . . Figure 10.6 . v~ H~ diagram & : 1995 -code model to the Cohen . levels of number ). ( Dehaene according triple processing

now turn to the more complex caseof the specificallyhuman ability to perform . exact calculations such as 2 + 3 = 5 or 12345679 x 9 = 111111111 is crucial of of levels that the distinction we believe Here again , organization to an understanding of the large network of areas , , including subcortical inferior parietal, and prefrontal components , that may yield different forms : Cohen, 1995). A coarsediagram of acalculiaif they are lesioned (Dehaene& of multiple circuits involved is shown in figure 10.6. At the simplest level, studies of nonnal adults have repeatedly suggested that most simple arithmetic problems , such as 2 x 3 = 6, are stored and retrieved from rote memory (for review, seeAshcraft, 1992). Anatomically, we believe that arithmetical fact retrieval rests on (a) left hemisphereperisylvian language areas for encoding the arithmetic problem into a verbal " " form such as two times three, and (b) a left corticostriatal loop for automatically " " retrieving the associatedverbal response( six ). This anatomical , may explain why fact retrieval deficits , while clearly speculative organization to encode digits into verbal form, are unable in that occur (a) patients : Dehaene such as in personswith pure alexia (e.g., Cohen & , 1995), or (b) in : DavidMcCusker Corbett the left basal lesions of with ,& , ( ganglia patients : : Denes son, 1988; Hittmair -Delazer , 1994; Whitaker, Habiger, & , Semenza ,& . Ivers, 1985)

247

De6cits : Acalculia andPrefrontal Dehaene : Changeux , & , Cohen

"

'

. . .

. . .

'

strategy choice and planning

semantic elaboration form visual number , " S J I ganglia thalan and basal left

At a second level, some arithmetic problems do not have a memorized answer and must be recoded in a different form before retrieval, a process " " that we call semanticelaboration (Dehaene& Cohen, 1995). For instance , evidence from normal subjects suggests that additions such as 9 + 6 are often recodedas 10 + 5 and that operandsmay have to be reorderedbefore retrieval, 2 + 7 being recoded as 7 + 2. Becausesuch recoding requires an understandingof quantitative properties (e.g ., that 9 is close to 10 or that 2 is smaller than 7), we think that it involves activation of the quantity representation in inferior parietal cortex. Indeed , left inferior parietal or parietooccipitotemporal lesions yield a classic acalculia of the Gerstmann type, which we claim is due to an inability to guide rote verbal memory retrieval . Such patients often show dramatic by quantitative knowledge of numbers with addition and impairments large multiplication facts that are least known and which therefore require semantic elaboration (e.g ., 8 x 7, 9 + 8). Subtraction and division problems , which are presumablynot stored in rote verbal memory, may also be disproportionately impaired. However, simple rote facts such as 2 + 2 or 3 x 3 are almost invariably spared (" .g . McCloskey, Harley, & Sokol, 1991), as predicted by their storing at a lower level. A third level is reachedfor operations that require holding intermediate results in working memory, for instance , when multiple digits or carry operations are involved. Researchwith normal subjects has demonstrated the involvement of visuospatial and auditory verbal working memory in complex calculation(e.g . Hitch, 1978; Logie, Gilhooly , & Wynn , 1994). Anatomically, verbal working memory results from the collective function of a network of areasconstituting the left supramarginalgyrus, the left inferior frontal region, the insula , and the left superior temporal gyrus (Paulesu , Frith, & Frackowiak . , 1993) Visuospatial working memory may rely on a right lateralized prefrontal parieto-occipital network (Jonides , Smith, Koeppe , et al., 1993; McCarthy , Blamire , Puce , et al., 1994). Finally, at a fourth level, the most complex arithmetic problems require es. Multidigit operations , for instance sequential planning and control process , involve the resolution, in a strict order, of many single-digit problems . The selection and execution of each elementary operation must be controlled and possibly corrected . Such higher-level supervisory functions are likely to involve an anterior neuronal network including the dorsolateral , 1988; Posner& prefrontal cortex and the anterior cingulate Cortex (Shallice Raichle , 1994). Some acalculicpatients show selective impairments in these , for instance , misapplying the addition procedure proceduralaspectsof calculation to a multidigit multiplication problem or failing to carry properly & McCloskey, 1987). Patientswith prefrontal lesions show deficits (Caramazza in nonroutine arithmetic word problems in which a novel sequenceof calculationsmust be planned to fit verbal instructions (e.g . Fasotti, Eling, & Bremer , 1992; Fasotti , Eling, & van Houtem, 1994; Luria, 1980). Frontallesioned patients are also impaired in a cognitive estimation test in which simple but unusual numerical questions must be solved (e.g ., what is the

248

Tests andClinical Syndromes Neuropsychological

, height of the highest building in London?, to which one patient responded " ". meters ) According to Shallice and Evans (1978), this test requires planning a searchfor relevant facts in memory and verifying the plausibility of a result. Both componentsmight be perturbed by frontal lesions . THE WISCONSIN CARD SORn N G TEST , PREFRONTAL CORTEX,

AND FAST RULE SWITCHING


While the large-scalenetwork of perisylvian, subcortical , parietal, and prefrontal areasthat collectively mediate the execution of complex arithmetical , it would clearly be premature to operations begins to be characterized undertakea detailed simulation of this entire network at the smaller scaleof individual neuronal circuits. Instead , we have again adopted the strategy of to model a only component of this network, namely, the role of attempting prefrontal cortex in working memory and task control. Two types of tasks -responsetasks were simulated : the delayed , in which an animal must memorize a responsecue throughout a time delay (Dehaene& : Changeux , 1991), and the Wisconsin Card Sorting test ( WCST in which have to discover ), subjects : , by trial and error, rules for sorting a deck of cards (Dehaene & , 1993). We focus here on the latter test, which is a classicneuropsychological Changeux test of prefrontal damage(e.g ., Milner , 1963). The problem facing our simulatednetwork is the samethat confronts subjects taking the WCST. A card bearing some colored geometric symbols is presentedat input, and there is a choice of four output stackson which to sort it . There are multiple available mappings between inputs and outputs, which correspond to the different ways of sorting the cards by color, by , and by number. Each time a card is sorted, the system is informed shape whether its responseis correct or not. The issue is, how can such a system find the correct sorting rule, and how can it switch rapidly to a new rule in casethe current one is found to be incorrect? In our simulation, we again solved the problem by introducing a layered . At the lowest level were input units coding for the hierarchicalarchitecture featuresof the input card, with separatenetworks for shape , color, and number . , and four output units coding for the four available sorting responses Above this level were layers of memory and intention units, isomorphic to the sensory and motor layers, but capableof maintaining a sustainedactivity , in the absenceof overt inputs and output, using recurrent excitatory . These units therefore provided the network with a " working connections " memory space in which to test anticipatory hypotheses about the coming . response At a higher level, above and beyond the basic mapping between memory and intention units, our simulation included a network of abstract rulecoding clusters(figure 10.7). These units were able to modulate an entire set of connectionsof the lower level. Their state of activity therefore drastically affectedinformation processing . When the cluster coding for the " color rule"

249

Dehaene : Acalculia : Changeux andPrefrontal Deficits , Cohen ,&

\.

Evaluation System

Reward Error cluster

.J
Auto -Evaluation Loop
t

Rul~ odlng clusters

Memory

Intention

Input

Output

of passing the Wisconsin CardSortingtest ( Dehaene & : Figure 10.7 Neuralnetworkcapable . 1991 ). Changeux was active , for instance, color , rather than number or shape information , was paid attention to and stored in the memory - intention circuit . The network then sorted the input cards by color . Changing the sorting rule simply consisted in changing the state of activity of rule- coding clusters. Because rules were encoded as patterns of activity rather than as fixed sets of connections , they could be changed rapidly under the control of a specialized evaluation system (see Agure 10.7). Whenever negative reward was received , signaling an erroneous response, an error coding unit Bred. This had the effect of desensitizing recently active connections among the rule- coding units , and therefore of destabilizing their current activity . As a result , the rule- coding cluster that was active when the error was made was deactivated , allowing the other sorting rules to enter a competition for the control of behavior . This process continued from trial to trial until a satisfactory rule was found , one that did not lead to a negative reward . An important property of this model is its ability to implement , in elementary form , a purely internal process of reasoning . When a negative reward is received , thus activating the error cluster , connections from intention to error clusters can learn the relation between the previous intention and the error . Hence, the previous intention is labeled as incorrect . Later , when a new rule is selected, it is applied to the memorized features of the

250

Tests andClinical Neuropsychological Syndromes

preceedinginput, creating a new pattern of activity of intention units. If, in fact, this pattern is identical to the previous one, this meansthat the network is about to make the sameerror a secondtime. The intention-to-error connections can recognizethis and avoid a secondfailure by automatically activating the error cluster , without waiting for an external reward input . This " " internal testing of rules , which we call autoevaluation , continues until a potentially valid rule is found, one for which it cannot be known a priori whether positive or negative reward will ensue . Several aspects of this model are relevant to the neuropsychology of , and intentional decision making. First, lesioning the planning, reasoning model leads to functional deficits similar to those seen in frontal-lesioned . Notably , damage to the rule-coding clusters or to the evaluation patients network yields perseverationssimilar to those seenin actual WCST performance , severalaspectsof the model are in rough (e.g ., Milner , 1963). Second . The correspondenceto actual anatomical and neurobiological structures memory and intention clusters may correspond to dorsolateral prefrontal cortex, where neurons maintaining a sustained activity during the delay -responsetasks and coding either for retrospective aspects period of delayed of past stimuli or for prospective aspects of future actions have been recorded (e.g . Fuster , 1997). The rule-coding network, with its ability to " " switch rules , , may correspond to the central executive (Baddeley rapidly " " " 1986), supervisory attentional system (Shall ice, 1988), or the attention for " action system (Posner & Dehaene , 1994) postulated by psychologists to underlie flexible task control, and anatomically related to the prefrontal and anterior cingulate cortex. Finally, the evaluation system may correspond to , 1988). Diffuse ascending neuroprefrontal-limbic loops (Goldman-Rakic modulatory systemsmay mediate the postulated effect of negative reward on prefrontal activity , perhaps via allosteric transitions of postsynaptic , 1991). Hence , speculative links receptor molecules (Dehaene& Changeux can begin to be drawn between high-level functions such as reasoning and the underlying anatomicaland even molecular structures .

CONCLUSION
Severaltypes of network models have been describedin this chapter . Some " " were framed at the global level of circuits of circuits and addressedthe large-scale interactions between brain regions involved in number processing and calculation , using the classic box-and-arrow diagrams of neuropsychology . Others aimed at providing a more detailed and necessarily more tentative simulation of the neuronal circuits underlying high-level -responsetasks or the Wisconsin Card Sorting behaviors such as the delayed test. Finally, a specific model of a single crucial circuit of the global model, the quantity representation , was simulated in detail using more plausible . We would like to biological assumptionsabout neuronal architecture conclude that a whole range of approach es to modeling may be useful in

251

Dehaene : Changeux : Acalculia and Prefrontal Deficits , Cohen ,&

neuropsychology . Much of cognitive neuropsychology is still dominated by functional box -and-arrow models , the usefulness of which is unquestionable for the initial charting of the main dissociations within a given cognitive domain . Nevertheless , detailed neuronal network simulations are also important , both for specifying the inner workings of the boxes in classic neuropsychological diagrams , and for drawing relations between behavioral and neurobiological data.

REFERENCES
Ashcraft : A reviewof dataandtheory , M. H. (1992 . Cognition ). Cognitivearithmetic , 44 , 15 . 106 . Oxford .: Oxford University , A. (1986 Press . ). Working , U. K Baddeley memory -Babic : Mehler - day , R., Bertoncini -old infantscategorize , J., & , J. (1991 Bijeljac ). How do four - 121 ? Developmental utterances . , 29 , 111 multisyllabic Psychology : Trabasso inferences andmemory in youngchildren , P. E., & . Nature , T. (1911 ). Transitive Bryant , 232 - 458 . , 456 : Clark view of cognitivenumber , J. I. D., & , J. M. (1988 ). An encoding Campbell complex processing : Comment on McCloskey : Goodman : , Sokol ,& (1986 ). Journal of Erpmmmtal Psychology - 214 General . , 117 , 204 : McCloskey of calculation es . In G. Deloche Caramazza , A,& , M. (1987 & : ) . Dissociations process - 234 X. Seron .), Mathematical Disabilities : A cognitive (Eds (pp. 221 ). neuropsychological perspective Hillsdale .: Erlbaum . , N.J
' neuronal . Paris : Fayard . . J. P. (1983). L homme Changeux : Dandtin, A (1916). Selectivestabilization of developing synapses . J. P., & as a mechanism Changeux for the specificationof neuronalnetworks. Nature , 264 , 105- 112. : Dehaene . J. P., & , S. (1989). Neuronal models of cognitive functions. Cognition , 33, Changeux 63- 109. : Danchin . J. P., Courrege , P., & , A . (1913). A theory of the epigenesisof neural networks Changeux . Proceedings Ct S , 70 , by selective stabilization of synapses of the National Academyof Scitn 2914- 2918. : Patte . In P. Marler & . J. P., Heidmann . T., & : H. , P. (1984). Learning by selection Changeux -Verlag. Terrace(Eds .), 11Itbiologyof learning(pp. 115- 139). Berlin: Springer Cohen : Dehaene , L., & , S. (1995). Number processingin pure alexia: The effect of hemispheric and task demands . Neuro Case , 1, 121- 131. asymmetries Cohen : Dehaene : Evidencefrom a case , L, & , S. (1996). Cerebralnetworks for numberprocessing of posterior callosallesion. Neuro Case , 2, 155- 114. : Verstichel Cohen , L , Dehaene , S., & , P. (1994). Number words and number non-words: A caseof to arabic numerals . Brain , 117 , 261- 219. dyslexia deep extending . G. (1984). Early number development : Discovering number spacewith addition and Cooper, R . In C. Sophian( Ed subtraction .), Originsof cognitive skills (pp. 151- 192). Hillsdale, NJ: Erlbamn . Corbett, A J., McCusker : Davidson . (1988). Acalculia following adominant , E. A , & , O. R , 43, 964- 966. hemispheresubcorticalinfarct. Archivesof Neurology Dehaene , S. (1992). Varieties of numericalabilities. Cognition , 44, 1- 42.

252

NeuropsychologicalTests and Clinical Syndromes

-related Dehaene of brainactivations in number , S. (1996 : Event ). The organization comparison and the additive factors methods . Neuroscience 8 47 68. , , Journal of Cognitive potentials Dehaene : Changeux modelof prefrontal cortexfunctionindelayed , 5., & , J. P. (1989 ). A simple - 261 tasks . Journal Neuroscience . , 1: 244 of Cognitive response Dehaene : Changeux CardSortingtest : Theoretical , 5., & , J. P. (1991 and ). TheWisconsin analysis in a neuronal network . Cerebral Cortex . ,1 , 62- 79 modelling Dehaene : Changeux of elementary numerical abilities : A neuronal , 5., & , J. P. (1993 ). Development . Journal model Neuroscience . , 5, 390- 407 of Cognitive Dehaene : Cohen calculation : A case , 5., & , L (1991 acalculia ). Two mental systems studyof severe with preserved . 29 1045 1074 . , , Neuropsychologia approximation - Neuropsychological Dehaene : Cohen mechanisms of subitizing andcounting , 5., & , L (1994 ). Dissociable evidence from simultanagnosic . Journal : Human of E .rperimental Psychology patients - 975 andPerformance . , 20 , 958 Perception Dehaene : Cohen an anatomical and functional modelof number , 5., & , L (1995 ). Towards processing - 120 . Mathematical . ,1 , 83 Cognition Dehaene : Mazoyer , 5., Tzourio , N., Frak , V., Raynaud , L , Cohen , L , Mehler , J., & , B. (1996 ). Cerebral activations number and : A PET . , during study Neuropsychologia multiplication comparison - 1106 34 . , 1097 ' . (1891 verbale avecagraphie . Mimoires ,J dela ). Surun casde cecite , suivi d autopsie Dejerine Societe deBiologie 3 201 . , , 197 ' - pathologique et cliniquedes different , J. (1892 es ). Contributiona l etudeanatomo Dejerine varietes dececite verbale . Mimoires dela Societe deBiologie , 4, 61- 90. Edelman brain . Cortical andthegroup -seledive , G. (1978 ). Themindful organization theory of higher brain . Cambridge , MA: MIT Press function Edelman Darwinism . New York : Basic Books . , G. (1987 ). Neural Fasotti : Bremer of arithmetical , L , Eling , J. J. C. B. (1992 ). The internalrepresentation , P. A. T. M., & : Frontalandposteriorpatients word problemsentences . BrainandCognition , compared - 263 20 . , 245 : van Houtem Fasotti of arithmetic , L , Eling word problems , J. (1994 ). Categorization , P. A T. M., & and nonnals frontal . Clinical and E , injured by of posterior patients Journal .rperimental - 733 . , 16 , 723 Neuropsychology cortex Fuster . New York : Raven Press . , J. M. (1997 , 3rd ed ). The prefrontal . (1992 Gallistel : Gelman andverbalcounting andcomputation . Cognition ,R , C. R., & ). Preverbal , 44 . , 43- 74 : Hillyard and speech of the right hemisphere , M. S., & , S. A (1971 ). Language Gazzaniga capacity - 280 . . Neuropsychologia , 9, 273 ., & : Smylie of language and cognition : Apsycho, M. S , C. E. (1984 ). Dissociation Gazzaniga two disconnected of . Brain 107 . 145 153 , , right hemispheres logicalprofile of fingeragnosia for right andleft, agraphia Gerstmann , J. (1940 and ). Syndrome , disorientation . Archives andPsychiatry acalculia - 408 . , 44 , 398 of Neurology -Rakic Goldman of prefrontal cortex . Trends in Neuroscience , P. S. (1984 ). Modularorganization ,7 , . 419 - 424 -Rakic of cognition Goldman : Parallel distributed in primate , P. S. (1988 networks ). Topography - 156 Review cortex . Annual association . , 11 , 137 of Neuroscience

253

: Acalculia and Prefrontal Deficits : Changeux Dehaene , Cohen , &

: Boller Grafman . J., Kampen , D., Rosenberg , A., & , F. (1989 ). Calallationabilitiesin a , J., Salazar - 194 . Behavioural . , 2, 183 Neurology patientwith a virtualleft hemispherectomy -termworkingmemory Hitch in mental arithmetic . Cognitive , G. J. (1978 ). Therole of short Psychology - 322 . , 10 , 302 -Delazer Hittmair : Denes andfads in calculation . Brain , M., Semenza , C., & , G. (1994 ). Concepts , - 728 117 715 . , ., Koeppe : Mintun , J., Smith , E. E , R. A , Awh , E., Minoshima , S., & Jonides , M. A (1993 ). Spatial in humans as PET . revealed Nature 363 623 625. , , by workingmemory : Wynn , K J., & , V. (1994 ). Countingon workingmemoryin arithmetic LogieR. H., Gilhooly - 410 . Memory andCognition . , 22 , 395 solving problem Luria cortical in man . Books , A R. (1980 , 2nded. New York: Basic ). The higher functions Macaruso : Aliminosa architecture of the cognitive , P., McCloskey , M., & , D. (1993 ). The functional -processing numerical : Evidence from a patientwith multipleimpairments . Cognitive system - 376 . , 10 , 341 Neuropsychology Mandler : Sheba : An analysis of its component es . ] ournRl , G., & , B. J. (1982 ). Subitizing of process : General , 111 , 1- 21. Erperimental Psychology -Rakic : , G., Blamire , A. M., Puce , A , Nobre , A C., Bloch , G., Hyder , F., Goldman , P., & McCarthy Shulman resonance of human cortexactivation , R. G. (1994 ). Functional magnetic imaging prefrontal task . Proceedings , 91 , of theNaHonal Academy of Sciences duringa spatial workingmemory . 8690 - 8694 : Basili in number mechanisms , M., Caramazza , A,& , A (1985 ). Cognitive McCloskey processing - 196 andcalculation : Evidence fromdyscalculia . Brain t:md . , 4, 171 Cognition : Sokol . M. (1991 of arithmetic factretrieval : An evaluation , M., Harley , W., & ,S ). Models McCloskey in light of findingsfrom normaland brain - damaged . Journal subjects of Experimental Psychology - 397 : Learning t:md . , Memory , 17 , 377 Cognition Nature - 695. , B. 0 ., Chalmers , M. (1977 , 267(June ). Are monkeys ), 694 McGonigle logical7 : Church . M. (1983 andtiming process . es Meck , W. H., & ,R ). A modecontrolmodelof counting Animal : Behavior Process es 9 320 . 334 , , Journal of Experimental Psychology Milner of brainlesions on cardsorting . Archives . , B. (1963 ). Effects , 9, 90- 100 of Neurology Mitchell :O ' Regan of a , R. W., Yao , P., Sherman , P. T., & , M. (1985 ). Oisaiminative responding truncatus rewarded stimuli . Journal ) to differentially , of Comparative dolphin(Tursiops Psychology - 225 . 99 , 218 Mount castle function : The unit moduleand , V. B. (1978 ). An organizing principlefor cerebral the distributed . In G. M. Edelman : V. B. Mount castle .), Themindful brain , & , (Eds , Cambridge system : MIT Press . , MA Paulesu : Frackowiak . J. (1993 correlates of the verbalcomponent , E., Frith , C. D., & , R. S ). Theneural - 345 . Nature of workingmemory . , 362 , 342 - 79 Posner : Dehaene networks . Trends in Neurosciences . , M. I., & , S. (1994 ). Attentional , 17 , 75 Posner : Raichle . NewYork: Scientific American . , M. I., & , M. E. (1994 ). Images of mind Ubrary Roland : Friberg of corticalareas activated . Journal , P. E., & ). Localization , L. (1985 by thinking of . - 1243 , 53 , 1219 Neurophysiology -Rumbaugh : Hegel in the chimpanzee , D. M., Savage , S., & , M. T. (1987 ) . Summation Rumbaugh - 115 . : Animal Behavior Process es . (Pan ) , 13 , 107 Journal troglodytes of & Psychology perimental

254

NeuropsychologicalTests and Clinical Syndromes

-Lorenz . , Reuter : Gazzaniga : , S. E , P. A , & , M . S. (1994). The disconnection syndrome Seymour . Brain basicfindings reaf6rmed , 117 , lOS- lIS . to mentalstrudurt. Cambridge shailice , T. (1988). From neuropsychology , UK. : Cambridge University . Press : Evans . (1978). The involvement of the frontal lobes in cognitive estimation Shallice . , T., & , M. E Cortez , 14 , 294- 303.

' : Adams of nmnerical . British transfonnations , C., & , N. (1987 ). Infants undentanding Sophian . , 5, 257 264 ] oumRl of Dtoelopmental Psvchoiogy : Cooper . G., Jr . (1980 of numbers infants . Scitna , P., & ,R ). Perception , 210 , Starkey by human - 1035 . 1033 . (1983 . S ., & : Gelman of intermodal numerical , P., Spelke ,E , R ). Detection Starkey correspon human infants . Science 222 179 181 . dences , , by . (1990 abstraction infants . Cognition : Gelman , R , P., Spelke , E. S., & ). Numerical by human Starkey - 127 . , 36 , 97 . F., Mayers . S ., Robert . T., & : Patterson ,K son , R , C. J. (1970 ,R ). Nmnbercodingin Thompson - 273 cortexof thecat . Science . assodation , 168 , 271 . (1991 von Fersen : Staddon inference , L., Wynne , C. D. L , Dellus , J. E. R , J. D., & ). "':ransitive - 1341 in pigeons . ] oumRl RlPsvchoiogy : AnilnQlBehlwior . fonnation Process a, 17 , 334 of Erperimenf . (1985 lesion . Neurology Whitaker : Ivers froma lenticular -caudate , H., Habiger ,R , , J., & ). Acalculia 35(Suppl . 1), 161 .

2SS

Dehaene : Changeux : Acalculia and Prefrontal De6cits , Cohen . &

11

Assessment of Neuropsychological Attention and ItsDisorders : Computational Models forNeglect , Extinction , and Sustained Attention
L. OwnbyandDylanG. Harwood Raymond

Patients who have undergone some form of neurological injury , most commonly that due to a cerebrovascular accident , may display a range of cognitive phenomena that suggest that they do not perceive parts of space around them . These phenomena may include such bizarre behavior as an inability to ' recognize the contralesional arm as the patient s own , failing to dress or shave the part of the body opposite the lesion , or an inability to attend to events occurring in the contralesional hemispace. Hemineglect may be manifest in more subtle forms , however , such as deficient performance in recognizing target letters in an array of letters placed in the contralesional hemispace, in correctly dividing a line into two equal parts , or in recognizing stimuli in the contralesional hemispace when another stimulus is simultaneously presented in the ipsilesional space. Hemineglect and related phenomena are often associated with damage to the parietal lobe , most often with dominant parietal damage and perhaps especially that occurring in the temporoparietal junction . Patients with this sort of damage seem to behave as though the half of space contralateral to the injury does not exist . This is true even though their sensory abilities are intact - these patients do not have, for example , hemianopia for the unattended hemispace. Such patients , when their attention is directed , can often correctly identify objects in the hemispace opposite the injury . When their performance depends on their own initiative in visually exploring space around them or on acting upon objects represented bilaterally , however , their performance suggests inattention to the area opposite their neurological lesion (see Rafal, 1997, for a more complete review of the cognitive manifestations of neglect ). The presence of neglect is usually most dramatic after acute focal injuries , but may persist after a considerable period of recovery . Neuropsychologists have devised a number of measures to assessneglect related phenomena . Two tasks that have been used to assessneglect are the line bisection and letter cancellation . In the line bisection task, the patient is asked to draw a line dividing a previously printed line in two . Patients who display neglect place the lines they draw to one side or another , away from the part of space they neglect . Rather than neatly cutting the line in two , for

t
f

. Figure 11.1 The line bisection task

A c

D s

0 k L ~ 0K doG G f 0

. Figure 11.2 The letter cancellationtask

, a patient with right parietal damageis likely to place the line far to example the right of the line' s true midline. This suggeststhat the patient does not notice, or experience . A similar phenomenon occurs , the left half of space with left (more properly, dominant) parietal damage , although less frequently . (See figure 11.1 for examples .) The reason for this difference is discussed below. In the letter cancellationtask (figure 11.2), the patient is presentedwith an array of various letters, often of various sizes and in various orientations. The patient's task is to detect all instancesof a target letter and to mark them. Patients who exhibit neglect will not notice the target letter in the . neglectedhemispace

258

NeuropsychologicalTestsand Clinical Syndromes

with hemineglect Various other sensoryphenomenaare associated , among them sensory extinction . Some patients who do not exhibit obvious sensory neglect may not recognize one of two stimuli presented to them simultaneously and bilaterally. For example , a patient touched lightly on either the or left arm may readily recognizeunilateral stimulation, but will recognize right one stimulus when touched bilaterally. The patient will fail to recognize only the stimulus presentedcontralateral to the parietal injury . The most commonly accepted interpretation of this phenomenon is that sensory extinction is a subtle form of neglect. Several mechanismshave been described to account for neglect and extinction, encompassingboth cognitive and anatomical explanations (see 1 : Valenstein Hellman, Watson, & , 1993, for a more complete review). Some have argued , for example , that deficits in line bisection or letter cancellation tasks result from lack of attention to the hemispacecontralateral to the injury , and perhapsmay be due to dysfunction of a mechanismby which the . Evidence patient disengageshis or her attention from the ipsilateral hemispace for and against this type of attentionalneglect has been presented , Fitzpatrick, et al., 1990). Others have argued that deficient (Coslett, Bowers on performances these tasks stem from a directional hypokinesia that becomes evident when the examiner demandsa motor performancefrom the . patient Similarly, evidence for and against this type of intentionalneglect also exists (Coslett et al., 1990). An explanation for extinction in the context of preserved sensation for unilateral stimuli is less obvious. One explanation that is intuitively appealing is that cortical or subcorticaldamagemay result in weakenedrepresentations of stimuli appearing in the hemispacecontralateral to the damaged . Even though the patient may be aware of such weakenedrepresentations hemisphere of stimuli when they are presented unilaterally, they may be below the threshold of awarenesswhen they compete with intact representations in the ipsilateral hemispace . These accountsof neglect are cognitive in that they rely on mental constructssuchas perceptualrepresentations for their explanation of neglect. Others have focused on the anatomy of . attention and its disorders 1 : Mesulam and his colleagues(Mesulam , 1981, 1990; Morecraft, Geula ,& a a for coherent anatomical account of Mesulam 1993 , , ), example present neural system for the control of directed attention. Many neglect-associated phenomenacan be understood as resulting from damageto various parts of this system . Mesulam argues that an anatomicalneural network comprising of , and cingulate cortices in addition to parts of the portions frontal, parietal basal ganglia and thalamus is responsible for the direction of attention in primates (figure 11.3). Eachanatomicalcomponent makesa specificcontribution to the diverse cognitive componentsunderlying attention. The parietal of space , and the frontal component component helps develop a representation " a Mesulam 1990 , , provides map for the distribution of ( ) suggests

259

: Harwood: Neglect , Extinction, and SustainedAttention Ownby &

of the anatomicalnetwork. Figure 11.3 Schematic

" orienting and exploratory movements. . . (p. 598). The portion of cingu" late cortex involved in the network provides a map for assigning value to " spatial coordinates (p. 598). Importantly, anatomical studies have shown that the organization of this network is consistentwith a parallel distributed , 1990; Morecraft et al., 1993). Subcomponents processingmodel (Mesulam of the network are internally cross -connectedin such a way that they function as independentbut interconnectedprocessingmodules . Its modular architecturemakesthis anatomicalnetwork especiallysuitable - the anatomical network can be for computational simulation represented a architecture that is similar in its structure and function. by computational -Schreiber Other investigators (e.g ., Cohen , Romero , Servan , et al., 1994) have used computational modeling to investigate attention-related phenomena . Cohen et at. (1994), for example , showed that a neural network for attention " could show evidenceof hemineglectwithout requiring a specific" disengage mechanism . These authors argued that the apparent disengage mechanism was an emergent property of the neural network. In this way, Cohen et at. showed that computational modeling can help in understanding attentionrelated phenomena . The anatomicalnetwork proposed by Mesulam and others also probably utilizes parallel distributed processing( Morecraftet al., 1993). It is thus reasonable to create an artificial network that could simulate the anatomical network' s activity using a similar computational architecture . Such a simulation would also allow study of the effects of various types of perceptual stimuli and lesions on the behavior of the network as well. In this manner , the effects of lesioning could be observed to detennine whether a network that simulates normal attentional process es would also simulate cognitive associated with lesions , such as deficits in line bisection or letter phenomena cancellationor perceptualextinction in the context of bilateral simultaneous stimulation. The purpose of these studies was thus to determine whether an artificial neural network could accurately simulate normal and disordered attentional process es in neglect-associated es. process

260

NeuropsychologicalTests and Clinical Syndromes

Figure11.4 Implementationof the network.

A secondarygoal was to show that the samenetwork architecturecould es seenin another disorder, attentionsimulatedisorderedattentional process deficit hyperactivity disorder. The purpose of this simulation was to determine whether the network model created to simulate one disorder could accurately simulate an unrelated disorder that presumably also depends on . dysfunction of the sameanatomicalstructures These studies should be regardedas preliminary since they involve a simulation of the anatomicalnetwork that is improbably simple. These preliminary studiesare presentedhere to illustrate the ways in which computational . modeling can facilitate understandingof neuropsychologicalphenomena

ADO NAt MODELS COMPUT es Study 1: Normal Attentional Process


Method The essentialcomponentsof the network include portions of the , and frontal cortices , as well as portions of the striatum. parietal, cingulate 11.4 illustrates the organization of the anatomicalnetwork asimple Figure mented in the artificial network. An input array of two setsof nine units with dimensions three units three units simulatesparietal cortex. The first three columns representthe left cortex, and the next three representthe right . The nine elementsin eachhalf of this array allow digital presentationof either an X or an 0 to either the right or left hemispace(see figure 11.4). An X can ' ' thus be encodedas l s representinga dark pixel and 0 s representinga light ' (imagine drawing two lines that connect the l s): 101 010 101 as illustrated in figure 11.4. The letter 0 can similarly be represented

261

: Harwood: Neglect, Extindion , and SustainedAttention Ownby &

The next two groups of input units representfrontal and cingulate cortices the two lower arrays of input units at the left of figure 11.4). Various patterns ( in the frontal units simulate different levels of exploratory activity thesechangeto representdifferent intensities of attention to the right or left . Cingulate units simulateincreasingmotivational levels demanded hemispace increased by exploratory behavior. A set of ten hidden units representsthe integrative function of the striatum and pulvinar; two output units allow a readout of whether the input pattern is an X or an 0 , whether it was presented in the right or left hemispace , and its distancefrom the midline. Outputs of 0.7, 0.8, and 0.9 representedan X at different distances from the midline, with larger values correlated with increasedinputs to the frontal units, while outputs of 0.3, 0.2, 0.1 in the same way represented O' s. An . output of 0.5 representedneutral attention, or a midline position in space The network was constructedand trained using a backpropagationalgorithm using available network simulation software (the Stuttgart Neural Network Simulator : Vogt , 1996). As learning in a backpropagation ; ZeIl, Mamier, & network is susceptibleto the effectsof local minima, a total of five networks was constructedand trained.
Results

A mean squarederror value of less than 0.001 was readily obtained over 4000 learning trials for each network. Comparison of network outputs showed that none differed significantly (all P ' s > .10). Results for one network are thus presented here. Comparison of network outputs before and after training showed that training changed outputs so that they were corred (table 11.1; compare columns under " Correct," "Untrained ," and
Table 11.1 Trainingsetandresults of lesioning Correct Left Untrained Trained 10%Lesion 20%Lesion 30%Lesion Right

Right Left

Right Left

Right Left

Right Left

Right Left

262

Tests and Clinical Syndromes Neuropsychol ~ icaJ

"Trained " . ' ) Inspection of the network s trained outputs indicates that the network readily learnedto discriminatewhether an X or an 0 was presented to it in either the right or left hernispace , and the distancefrom the midline at which it was presented . This first study thus conArmed that the general strategy of creating a backpropagationnetwork that would simulate normal attentional process es could be successful . It thus was warranted to proceed to the next phaseof developing the model, simulating neglect.
.: Simulation Study 1 of

Heminegled

Method Experimentallesions of artificial neural networks have been used to study the effects of neurological injuries in clinically studied disorders ' such as dyslexia (Hinton & Shallice , 1991) and Alzheimer s disease( Tippett & Farah , 1994). Units, or connectionsbetween units, may be removed from the network to simulate the effects of damageto the simulated underlying neural structures . Theselesionsthus presumablyrepresentdamageto specific areas of cortex or to the fiber tracts connecting these structures . In this study, different numbers of connections between input and hidden units were deleted to create different levels of injury severity. Connections were deleted from primarily the right - or left-sided input units to simulate lateralized neurological injury . Connections were randomly deleted from the entire network to simulatethe effectsof diffuse injury , to assess the effectsof this type of injury on network performance . It was hypothesizedthat unilateral , while diffuse injury would reduce injuries would simulate hemineglect network accuracy without affecting in which hemispacestimuli were perceived . As before, a total of five networks were constructed , trained, and tested in each condition to assure that findings from any single network were not due to local minima achieved in the learning process (Ownby , 1995; 1996a ). Results Resultsdid not differ significantly acrossnetworks, and results for one representativenetwork are thus presentedin table 11.1. Resultsof right , left, and diffuse experimentallesions were consistent with the hypothesized relations ; Table 11.1 illustrates simulated right hemisphereinjury . Correct network responses are listed in the first column for comparison . With experimental lesions generally, the performance of the network slowly became less accuratewith lesions of increasing severity. At more severe levels of unilateral injury , the network' s performancesuggesteda tendency to report that stimuli were present in the hemispacecontralateral to the injury , consistent with clinically observedhemineglect . Inspection of the "Left " column, " " for example , under 20% lesion shows that the accuracyof the leftward , while the magnitude of the responsein the right output responsewas reduced -Ranks test; was significantly increased( Wilcoxon Matched-Paris Signed z = - 2.20; P = .OJ and increased ). Decreased magnitudeof leftward response

263

: Harwood: Negled , Extinction , and SustainedAttention Ownby &

- lesioned network is magnitude of rightward responsein a right hemisphere thus consistent with negled . It may be recalled that patients with right hemispherelesions are expeded to neglect the left side of space and to demonstrate increased attention to the right hemispace . Left hemisphere lesionsproducedsimilar results . Diffuse lesions , as hypothesized , reducedthe network' s accuracy . The network based , but did not produce lateralization ' on Mesulam s anatomical model thus, at least at this rudimentary level, simulatedboth normal and disorderedattentional process es.

Study 3: Simulation l of Attention -Deficit Hyperactivity Disorder


Attention -deficit hyperactivity disorder (ADHD ) is a psychiatric condition most often diagnosed among children, who may present evidence of disordered attention, motivation, and behavioral or emotional regulation. A ' hallmark of the disorder , however, is the patient s difficulties in sustaining attention so that he or she is described as unable to pay attention in the classroomor to completeother activities that require sustainedattention. Patients with this disorder often also have difficulties in sustaining attention on the continuous performancetest (CPT ) . This measurerequiresthat a , patient monitor some form of display on which test stimuli are presented and makea responsewhen a target stimulus appears . Most often, the patient is required to differentiate relevant from irrelevant stimuli, for example , by instances of the letter r in a serial of other letters. detecting presentation Difficulties in sustainingattention are also observed in patients who have sustainedneurological insults. Although attentional problemsmay be seenin patients with focal injuries, they are also seen in patients who have undergone , as may occur in patients with closed relatively mild but diffuse trauma head injuries (Gronwall, 1989). It thus seemslikely that a disorder of the network for directed attention might underlie the sorts of attentional deficits observedin clinical populations. This problem thus could be investigated by simulation of normal and disorderedperformanceon a measureof sustained attention, the CPT. Different types of lesionsmight be investigated with the what types and foci of lesions might causedeficits in purpose of assessing on a simulation of the CPT. performance Method The network describedabove was trained and then subjectedto random lesionsof varying degreesof severity. In this instance , however, the network was trained to respond with 0.1 for all letters other than X, for which it gave an output of 0.9. lesions were not lateralizedand stimuli were ' presented bilaterally (Ownby , 1996b). The network s performance was assessed before and after training and after eachstageof lesioning.
Results Results show that minor degrees of lesioning did not substantially ' affect the network s performance . lesions of greater extent , however , resulted

264

NeuropsychologicalTests and Clinical Syndromes

in substantial decrementsin performance . Inspection of the network' s outputs showed that although accuracy declined with increasing severity of lesion, outputs did not suggest increasedor decreasedattention to either . The network thus accurately simulated one aspect of patients' hemispace accuracyin detecting a target stimulus. performanceon the CPT decreased Study 4: Simulation of Line Bisection and Letter Cancellation As reviewed above , several authors have suggestedthat distinct behavioral types of neglect behavior may occur and that these subtypes are related to the locus of the underlying neurological injury . For example , frontal injury is said to impair letter cancellationperformancemore than line bisection, since the patient with this type of lesion may have a deficit in the exploratory behavior required for performance of this task. Conversely , patients with worse on line bisection tasks , injury may parietal perform perhaps because are more in the abilities they relatively impaired visuospatial required for on this task. The performance purpose of this study was to investigate whether an artificial neural network constructed to simulate the different contributions of parietal and frontal cortex to performance on these tasks could do so (Ownby , 1996a ). Method A three-layer backpropagation network was again constructed , with the input layer representing the three cortical elements of the corresponding anatomicalnetwork as before. The middle, or hidden, layer, represented the intersection of inputs and outputs of the cortical units in the striatum and pulvinar. Two output units representedeither line bisection or letter cancellationoutputs. The network was trained to an acceptable level of less than a mean error of 0.001 one unit was ( ); accuracy squared output trained to make judgments about the position of letters at three positions either to the right or left of midline, while the other output unit made similar judgments in a line bisection task. The organization of training data made the letter cancellationtask more dependenton frontal inputs, while the line bisection task was more dependent on parietal inputs. The network was then experimentally lesioned in several ways, including presenting it with degraded information and by removing connectionsbetween layers. These methods representdamageto cortical areasor to fiber tracts. Results After training, the network was able to perform both letter cancellation and line bisection tasks at a 100 percent correct level of accuracy . " " Lesions resulted in degraded network performance , with frontal lesions producing greater decrementsin network performanceon letter cancellation " " tasks , whereas parietal lesions produced greater decrementsin line bisection task performance . Lesions , however, produced some decrementsin performance in the ipsilateral as well as contralateralhemispace on both tasks .

265

: Harwood: Neglect, Extinction, and SustainedAttention Ownby &

Study 5: Simulation

of Extinction

Method

A three-layer network was constructed employing acounter : Palmer , 1991). Using this architecture propagation architecture (Hertz, Krogh, & the numerical data coded to representeither an , input layer accepted X or 0 presentedin either the left or right visual or tactile field. Employing an unsupervised , or Kohonen , learning algorithm, this layer and its connections with the next representsa pattern recognition device. The middle, or hidden, layer, fed pattern recognition information forward to four output units. These units were trained with a supervisedlearning algorithm to produce different numerical values when input sensory units were presented with digitized versions of an X, 0 , or no pattern (the processing layer). These layers thus functioned as a cognitive processing layer, accepting information about sensorypatterns presentedto the network and producing reducedoutput about the nature and position of those patterns. The network was created and then trained as described above to a low level of error (mean squarederror < 0.001). The network was trained with examplesusing only one stimulus to one sensoryfield on any single learning trial. The network was then experimentally lesioned in severalways to simulate damageto fiber tracts or cortical areasas might occur in stroke. The at different levels of lesioning (10, 25, or 50 percent of network was assessed units or links) and with lesions in the areassimulating the right or left hemisphere . Both before and after lesioning, the network was presentedwith test stimuli simultaneouslyto both sensory fields to simulate its behavior under conditions of simultaneousbilateral stimulation. Outputs were recorded to assess the effectsof site and extent of lesions(Ownby , 1997a , b) Results The network was readily able to distinguish the two letters in . Observation of the network showedthat when the hidden sensoryhemifields the samenumber of elementsas patterns contained , one element in the layer hidden layer was active for each distinct pattern. This result suggestedthat the pattern recognition portion of the network was appropriately trained. . For Outputs from the output layer after training were 100 percent accurate an X in when information for the left hemifield was , pattern presented example to it , the network consistently reported that an X was presentedto the correct hemifield. This was true for all other possible combinations of pattern inputs and outputs. Presentationof bilateral inputs to the network . Rather than produce two correct outputs when reduced network accuracy with two stimuli , the network produced one correct output for presented eachpattern, resulting in 50 percent accuracy . in both and Lesioning pattern recognition supervised learning layers results consistent with inattention . Removal of small produced phenomena numbers of links (e.g ., 10 percent had little effect on network outputs. ) in both Removalof larger numbers , however, reducednetwork performances and simultaneous bilateral conditions. Bilateral simulta single presentation

266

Syndromes NeuropsychologicalTests and Clinical

neous stimulation of lesioned networks showed that they were relatively more sensitive to the effects of this condition; their performance was degraded (to less than 50 percent correct) before lesions produced performance decrementsin the samenetworks to unilateral presentation . Removal of larger numbers of links produced performancedecrementsin both conditions . As in the previous study, removal of large portions of the network on either side produced network outputs consistent with the clinical phenomenon of hemineglect . lesions of the perceptual recognition layer were more likely to result in errors suggesting inattention of the contralateral , while errors in the processinglayer resulted in both contralateral hemispace perceptualerrors but also misrecognitions(e.g ., reporting a 0 for an X ).

DISCUSSION
Theseresults show that a computational model basedon the modular parallel distributed processing architecture of the anatomical network proposed by Mesulam can accurately simulate normal attention as well as some . These simple models demonstrate that a computa aspectsof its disorders tional model can thus accurately simulate directed attention to either hemispace at different levels of exploratory effort and with different motivational . It also shows how computational modeling can be used to investigate saliences . Suchmodeling can thus shed light on neuropsychologicalphenomena the mechanismsaffecting cognitive test performance in normal and disordered -Schreiber , Servan , Targetal ., 1992, for example , patients (seeCohen for a discussionof modeling aspectsof neuropsychologicaltest performance ' in schizophrenia ). The models successlends at least indirect support for ' Mesulam and colleagues proposed anatomical network that may underlie directed attention. Enthusiasmfor the success es of these elementary models in simulating attention, however, should be temperedby considerationof the things about attention that they do not simulate . For example , although the model simulates some aspectsof normal and disordered performance on the CPT, it -positive responses(when the person tested erroneously does not generate false respondsthat the target is there when some other stimulus is present . The ) present model simulatesreduced accuracyin detecting the target -negatives(one aspectof disorderedperformanceon the Cpn , stimulus , false but cannot reproduce the clinically observed converse . It can be speculated -positive responses that impulsive false be may generatedby the interaction of a responding system that is unsuccess fully regulated by a detection and control system . Further developmentof the attentional model presentedhere -positive responses in ADHD . might provide an improved simulation of false Another important aspectof attentional phenomenaand its disorders that is not adequately explained by this model is the relatively greater importance of nondominant (usually, right ) parietal cortex in the development of hemisensory neglect. It is widely assumedthat this effect is the result of

267

: Harwood: Neglect, Extinction, and Sustained Attention Ownby &

nondominant parietal cortex specialization for the perception and integration of spatial information . The present model implicitly assumes that the function of both right and left parietal cortex is the same; this assumption is clearly incorrect . Again , further development of the model may more adequately simulate the function of the anatomical network . Yet another consideration in evaluating this model is the extent to which the network architecture chosen for it is appropriate . It may be more appropriate to simulate the parietal cortex element of the network by using a network architecture that is more suitable to simulation of the perceptual and integrative role of parietal cortex . As demonstrated in study 5 (simulation of sensory suppression or extinction ), a self- organizing or Kohonen input layer (see Hertz et al. for a more detailed explanation ) simulated the perceptual function implicit in the functioning of the parietal cortex . This sort of hybrid network may thus more accurately simulate the different functions of the different modules composing the network . Given these limitations , it may seem surprising that this model works as well as it does in reproducing at least some of the phenomena associated with normal and disordered attention . As has been seen, it does a fairly good job at some things , and demonstrates how computational modeling can be used to address questions about neuropsychological phenomena . Further ' development of the model may address some of the model s failings and allow it to more accurately reproduce attention - related phenomena . This development is in progress .

REFERENCES
-Schreiber Cohen : Farah , j . D., Romero , R. D., Servan of spatial , D., & , M. j . (1994 ). Mechanisms attention : Therelation of macrostructure to microstructure in parietal . Jou ~ l of Cognitive negled - 387 Neuroscience . , 6, 377 -Sdtreiber Cohen : Spiegel : , j . D., Servan , D., Targ , D. (1992 ). Thefabricof thoughtdisorder , E., & A cognitive neuroscience to disturbances in the of context in . approach processing SChizophrenia In D. Stein& : j . Young(Eds .), Cognitive science andclinical disorders : (pp. 99- 126 ). New York Academic Press . Coslett : Hellman , H. B., Bowers , D., Fitzpatrick , E., Haws , B., & , K. M. (1990 ). Directional hypo - 486 kinesia andhemispatial inattention in negled . Brain . , 113 , 475 Gronwall andpersisting effedsof concussion on attention andcognition . , D. (1989 ). Cumulative In H. S . Levin & : A L. Benton . Mild head . 153 162 . New , H. M. Eisenberg (Eds ), ( ) injury pp , York : OxfordUniversityPress . : Valenstein . In K disorders . M. Hellman , K. M., Watson , R. T., & , E. (1993 ). Neglectandrelated Hellman & : E. Valenstein .), Clinical .). New York : Oxford University (Eds , (3rd ed neuropsychology Press . : Palmer to the theory Hon . Hertz , j ., Krogh , A, & , R. G. (1991 ). Introduction of neural computa
Reading , MA : Addison-Wesley. Hinton , G. E., & : Shallice : Investigations of acquired , T. (1991). Lesioning an attrador network J Review C R , 98, 74- 95. dyslexia. Psychoiogi

268

Tests andClinicalSyndromes Neuropsychological

networkfor directed Mesulam . M-M. (1981 attention andunilateral . AnM Isof ) A cortical neglect - 325 . , 10 , 309 Neurology -scaleneurocognitive Mesulam , M-M. (1990 networksand distributedprocessing ). Large for attention . Annals , language , andmemory - 613 . , 28 , 591 of Neurology Morecraft . J., Geula : Mesulam ,R of co , C., & , M-M. (1993 ). Architecture Mectivity within a -parietal -frontal networkfor directed attention . Archives , 50 , cingula neurocognitive of Neurology . 219 - 284 -related . L (1995 modelof stroke : Preliminary development ). A computational Ownby R hemineglect . Presented at the Workshop on NeuralModelingof Cognitiveand BrainDisorders , Studies , Universityof Maryland Universityof MarylandInstitutefor Advanced , Computing Park . MD. College -related . L (l996a modelof stroke ,R ). A computational Ownby (abstract ). Journal of hemineglect theIntern Rtional , 9, 38. Society Neuropsychological model of attention , R. L (l996b). A computational - defidt Ownby / hyperactivitydisorder Rtional (abstract ). Journal , 9, 24. of theIntern Neuropsychological Society . L (1991a networkmodelof sensory ,R . Presented at the annual ). A neural Ownby suppression of the International . 1991 , Orlando , FL , February meeting Society Neuropsychological afterstroke : A simulation . Presented , R. L (1991b at the annual meeting ). Sensory Ownby neglect of the Cognitive Neurosdence . , Boston , March1991 Society Rafal : Cognitive , R. D. (1991 . In TE . Feinberg ). Hemispatial neglect neuropsychological aspects and M. J. Farah(Eds .), Behavioral and neuropsychology - 335 : (pp. 319 ). New York neurology -Hill. McGraw : Farah . (1994 modelof namingin Alzheimer , L J., & , M. J ' s disease : ). A computational Tippett ?Neuropsychology . , 8, 3- 13 Unitaryor multipleimpairments Zell, A., Mamier : Vogt, M. (1996 , G., & ). StuttgartNeuralNetwork Simulator , usermanual , version 4.1. Stuttgart : Institute for Paralleland DistributedHigh Performance , Germany . [ TheWorld Wide Web pageon the simulatoris available , Universityof Stuttgart Systems at: http:// www.inEormatik . unistuttgart .de is available / ipvr/ bv/ projekte / snns /sMs.html. Thesoftware for download at: ftp:// ftp.informatik .unistuttgart .de .] / pub / SNNS

269

: Harwood: Neglect , Extindion , and SustainedAttention Ownby &

12

The Neural Basis of Lexical Retrieval


Daniel Tranel , Damasio

Hanna Damasio , and Antonio R.

In neuropsychology , there is a long tradition of studies concernedwith the of which neural structures support the process whereby, in the question courseof speaking , writing, or thinking, we retrieve words denoting various entities, actions , For many decades , and relationships , the common answerto this question has focused on left hemispherestructures in and around the , This focus , and in particular, on the Wernicke and Broca areas sylvian Assure in derives large measure &om studies in brain- damaged patients with , An , in whom word -retrieval (naming) defects are highly &equent aphasia in this answer is that the classic areas are activated assumption language directly by nonlanguageareaswhich support the conceptsfor which pertinent words are being retrieved; so, for example , the process of perceiving and recognizing a familiar face (" my mother" ), supported by structures in , would in turn activate word -retrieval occipital and occipitotemporal cortices structuresin the perisylvian language areasthat would generate the appropriate " name ( Virginia " ), Based on new findings &om lesion studies (H, Damasio : Damasio , Grabowski, Tranel, et al" 1996; Tranel, Damasio , & , 1997a ) and functional imaging studies (H, Damasio et al" 1996; Martin , , et al" 1995; Mazoyer, Tzourio, Frak , et al" 1993; Petersen , Haxby, Lalonde Fox, Posner et al" 1988), as well as related work using electrophysiological , : Uematsu , 1995; Lesser , Gordon, & , 1994; Nobre, Allison techniques(Dehaene : McCarthy , 1994; Nobre & : McCarthy , 1995; Ojemann, 1991), we , & have proposed an alternative answer , we have , Acally Speci proposed that although perisylvian structures (including early auditory and somatomotor cortices ) are indeed involved in the transient reconstruction and explicit phonemic representation of word forms, additional neural sites mediate between the structures that support , These mediational conceptual knowledge, and the perisylvian structures structures thus trigger and conduct the processof word reconstruction (H, Damasio et al" 1996; Tranel et al" 1997a ), Our account incorporates the notion that the retrieval of word forms dependson the transient reactivation of the phonemic and morphologic structure of given words within the , visual) and appropriate early sensorycortices (e,g " auditory, somatosensory motor -related structures (A , R, Damasio : Damasio , 1990; A , R, Damasio & ,

1992); the traditional Wernicke and Broca areasare partly contained within suchsensoryand motor structures . However, we propose that the phonemic morphologic reactivation pertaining to a given word is directed from a , but largely variety of regions located in higher-order associationcortices outsidethe early sensory and motor sites alluded to above , and, consequently . Those regions operate , largely outside the traditional languageareas " as third -party" neural mediators between , on the one hand, the regions that support conceptualknowledge (which are distributed over varied association cortices ), and, on the other, the sensorimotor regions in which the -morphologic structure can be transiently reconstructedduring the phonemic word -recall process(e.g ., naming), or instantiated during the perception of a word (e.g ., reading ). We believe that for most persons , these regions are located in the left , and are used as lexicalmediationunits (H. Damasio et al., 1996). hemisphere Once activated by the evocation of a given concept , these units, which do -morphologic information in explicit form, promote the not contain phonemic activation of the linguistic information necessary to reconstructa word form, momenta rily , in sensorimotor terms. These units also promote activation of syntactic information necessaryfor the proper placementof a word in the . In schematicform then, this is a tripartite phrasesand sentencesbeing planned model in which mediational structures interlink between conceptual and implementation structures , in two directions: from conceptualstructure to the word -form reconstruction required for language acts , and in reverse , from the perception of a word toward its usual conceptual structure corre. Our account is compatible with levelt ' s (1989, 1992) proposal spondences " of lexical mediation units termed " lemmas , basedon cognitive experiments in normal subjects , which has also been applied to retrieval of words for actions (Roelofs , 1993). The general theoretical background for this proposal is the framework for synchronousretroactivation from convergenceand divergence zones (A . R. Damasio t Damasio , 1989a , b; A . R. Damasio& , 1994; seealso Kosslyn, 1994). The framework posits that retrieval of conceptsdependson the reconstruction of imagesor actions pertaining to characteristics and properties of entities . The imagesand movementsthemselvesare reconstructedtransiently in sectors of early sensory and motor structures . However, the reconstruction is directed from separate system components in higher-order association cortices , which contain dynamic regions which interlock feedforward and -divergence feedback projection neurons . Those regions , the convergence zones , thus hold a preferentialbut probability driven and modifiable dispositional capacity to signal directly or intermediately to sensory or motor . regions, whenever the zone receivesappropriate signals Below, we review new work from our laboratory, which has dealt with several specific issuesconcerning word retrieval and the lexical mediation is whether the neural systemselaboratedabove. The first issuewe addressed on which retrieval systems conceptual knowledge depends are separable

272

NeuropsychologicalTests and Clinical Syndromes

from the neural systems on which word retrieval depends. Also , we tested the idea that lexical mediation systems would be located in the left hemisphere , and in particular , the question of whether the retrieval of words denoting concrete entities belonging to different conceptual categories would depend on relatively separable regions in higher - order cortices in the left temporal lobe . We also explored the hypothesis that different types of words - for instance, words for entities (nouns ) vs. words for actions (verbs )- would depend on separate neural regions . A few comments on terminology are in order . We use the designation concreteentities to refer to persons, animals , fruits and vegetables , tools and utensils , and other items that belong to varied conceptual classes , and that can be designated by an appropriately specific word (a proper or common noun ). Concrete entities are mapped in the brain at different levels of contextual complexity. This permits us to classify a given entity along adimension that ranges from unique (an entity belonging to a class with N = 1 and depending on a highly complex context for its definition ), to varied nonunique levels (entities processed as belonging to classes with N > 1, having members whose definition depends on less complex contexts ). Also , many as alluded to above , it is important to distinguish between retrieval of conceptual knowledge about entities , on the one hand , which relates to the " " traditional term recognition (i.e., knowing what an item is), and retrieval of the word fonns for entities , which refers to the traditional term naming. In our experiments , we required subjects to recognize and to name unique entities (persons) at a subordinate level , and to recognize and to name non unique entities at a basic object level (cf. Rosch, Mervis , Gray , et al., 1976 ). The hypotheses were approached by studying patients with neurological lesions in various components of the large -scale neural system presumed to be related to the language lexicon . The demonstration of a defect in the retrieval of words relative to particular categories of entities was used as evidence for the relation between the putative system and the access to the lexicon for that particular conceptual category in the normal human brain . We also mention briefly some of our new work with functional imaging (positron emission tomography , PET).

'AL OFWORDS RETR ~ \: I FORNONUNIQUE CONCRETE ENTITIFS


Background

For over a decade , investigators in severallaboratorieshave noted intriguing dissociationsin naming ability in patients with and without aphasia(A . R. Damasio : Van Hoesen : McCarthy , 1983; , Damasio , 81 , 1982; Warrington 81 Hart Berndt, 81 : Caramazza , 1985; Pietrini, Nertempi, Vaglia, et al., 1988; Basso : Laiacona , Capitani, 81 : , 1988; A . R. Damasio , 1990; Franklin , Howard, 81 Patterson 1995 81 : Budin 1988 , ; Goodglass , ; Goodglass , Wingfield, Hyde, et al., 1986; see also Small , Hart, Nguyen, et al., 1995). These observations

273

Tranel 1 : Darnasio : TheNeuralBasis of Lexical Retrieval . Darnasio .&

suggestedthat, following a lesion, naming of items from various conceptual . The possibility categorieswas not equally compromised across categories was raised that there were different neural systemsrequired for the retrieval of words for different classes of concepts . One frequent finding has been that brain- damagedpatients demonstrate ) entities, as compared greater impairment in naming living (natural, animate to artifactual (nonliving , manmade , inanimate ) ones (Bassoet al., 1988; Hart & : Gordon, 1992; Hillis & : Caramazza : Job, , 1991; Pietrini et al., 1988; Satori & : Gainotti, 1988). In a few cases 1988; Silveri & , the defect for living entities has been reported as being disproportionately severe for one particular , it has been found that either the category of animals category- for instance : Gordon, 1992) or that of fruits and vegetables (Hart et al., 1985; (Hart & Farah & : Wallace . In some cases , 1992) was relatively more affected , such - that have in been connection with name patterns reported comprehension is, patients had disproportionate impairment in the comprehensionof names of living entities, comparedto comprehensionof namesof artifacts (McCar: Warrington, 1988; Warrington & : McCarthy , 1987). A few caseshave thy & been reported in which the dissociation occurred in the other directionthat is, artifacts were disproportionately impaired , and living entities were : Caramazza : Humphreys , 1991; Sacchett& , 1992; relatively spared (Hillis & : Farah : McCarthy , 1983). Other , & , 1996; Warrington & Tippett , Glosser , disproportionate patterns of defects have also been reported, for example of name for : Budin, body parts (Goodglass & comprehension impairment 1988), or relatively sparednaming of body parts (Goodglasset al., 1986). Dissociations between the ability to processabstract words vs. concrete words have also been reported. For example , patients may lose the ability to processconcrete words, but remain capableof processingabstract words : Tranel, 1990; A . R. Damasio , Damasio , Tranel, et al., 1990; (A . R. Damasio& : Shall ice, 1984). The reverse pattern Warrington, 1975, 1981; Warrington & has also been reported, that is, a patient whose defect in word retrieval was " " especially marked for abstract words, and who was less anomic for concrete words (Franklin et al., 1995). -driven, and most consisted Few of the studies cited above were hypothesis of single-case reports. There was often incomplete sampling of the relevant categories . Another factor which hinders the interpretation of is the inconsistent use of naming , on the one hand, and previous findings , on the other. These capacitiesare quite separable , and they are recognition in brain Tranel dissociated see : , Damasio , & ( damaged patients frequently Damasio , 1997b , some investigators have used the two concepts ). Nonetheless interchange , ably (e.g., Martin , Wiggs, Ungerleider, et al., 1996; Farah : McMullen , 1996; Riddoch & : Humphreys : Hum, 1987; Sacchett& Meyer, & phreys, 1992), and it is therefore difficult to understand what was actually wrong with the patient or what process was actually being studied. (It should be acknowledged , however, that in normal subjects , recognition and

274

Tests andClinical Syndromes Neuropsychological

, since the naming of an item that is naming may be difficult to dissociate shown for recognition cannot be voluntarily suppressed .) The most important limitation in the extant literature as far as the neural basis of lexical retrieval is concerned , is that neuroanatomicaldata are quite limited. In most cases , the neuroanatomicalstatus of the patient was mentioned . , and the lesions were not analyzed systematically only in passing were often because shared some Subjects they grouped together aspect of their cognitive pro61es , regardlessof their lesion status , and inspection of such groups frequently revealsthe subjectsto have lesions in many different areasof the brain, even different hemispheres . We conducteda systematicstudy regarding word retrieval (A . R. Damasio et al., 1990; see also A . R. Damasio , 1990), in which the idea that there are separateneural systems involved in the naming performance for different lexical categorieswas supported . In another study, we were able to obtain evidence for selective involvement of left temporal cortices in postmortem a case of progressive loss of word retrieval for concrete entities (GraffRadford , Damasio , Hyman, et al., 1990). Recentstudies from our laboratory have confirmed and extended the initial findings (H. Damasio et al., 1996; Tranel, Logan, Frank , et al., 1997).
New Lesion

Studies

We have added to previous findings on the breakdown of retrieval of words for nonunique entities by pursuing systematicallytheir possibleneural correlates , 1992; A . R. Damasio (A . R. Damasio , Brandt, Tranel, et al., 1991; A . R. Damasio & Damasio , 1992; A . R. Damasio , Damasio , Tranel, et al., 1990; A . R. Damasio & Tranel, 1990 , 1993; H. Damasio et al., 1996; Tranel, 1991; Tranel et al., 1997a ).
In the studies summarized below , we tested the following hypotheses : ( 1) retrieval of words denoting concrete entities depends on neural structures that are anatomically distinct from the neural structures on which concept retrieval (recognition ) for the same entities depends; (2 ) retrieval of words related to separate conceptual categories of concrete entities (animals, fruits and vegetables , tools and utensils ) is mediated by distinct neural systems .

Subjects Brain - Damaged Subjects To address the hypotheses stated above , we conductedvisual recognition and naming experimentsin 127 brain-damaged . Based on measurementwith the subjects with single, unilateral lesions -Old Aeld questionnaire Geschwind , 115 subjects were right -handed , 6 were left-handed and 6 had . mixed handedness The , subjectswere selectedfrom ' the division s patient registry so that as a group, they would permit us to . The subjectshad lesions located in the left samplethe entire telencephalon or right hemisphere , and in varied regions of the cerebralcortex, causedby

275

Tranel, Damasio : Damasio : The Neural Basisof Lexical Retrieval ,&

either cerebrovascular disease (n = 106 ), herpes simplex encephalitis (n = 5), or temporal lobectomy (n = 16). All had I Qs in the average range or higher ; had a high -school education or higher ; had been extensively characterized neuropsychologically and neuroanatomically ; and had no difficulty with the attentive inspection of visual stimuli . None of the subjects had severe aphasia at the time of these experiments . Control Subjects Nonnal control subjects (n = 55 ) were drawn from our visual recognition and naming studies described elsewhere (A . R. Damasio et al., 1990 ). They were matched to the brain -damaged subjects on age, education , and gender distribution . (Gender - related effects on visual recognition and naming are generally of fairly small magnitude [e.g ., McKenna & Parry , 1994 ], with the most consistent finding being that women are better than men at naming fruits and vegetables , and men are better than women at naming animals. Hence, we used proportionate numbers of men and women in the brain -damaged and control groups , rather than analyzing the data separately by gender .)

Stimuli The stimuli for the study were 300 nonunique entities, comprising 161 of the black-and-white line drawings from the Snodgrassand Vanderwart set (1980), and 139 additional black-and-white and color photographs. Five categories are representedin the stimulus set: animals (n = 90), fruits and vegetables(n = 67), tools and utensils (n = 104), vehicles (n = 23), and musical instruments (n = 16). Here, we report results for the animals , fruits and vegetables . , and tools and utensils categories
Procedure The entities were depicted on slides and shown in random order one by one on a Caramate 4000 slide projector , in free field . For each, the subject was asked to tell the experimenter what the entity is (' What is " this? ). If the subject produced a vague or superordinate -level response (e.g ., " " Some kind of animal ), the subject was prompted to " Be more specific; tell " me exactly what you think that thing is. Time limits were not imposed . All responses were audiotaped . Data Quantification For each stimulus , the response Neuropsychological of each brain -damaged patient was scored as correct if it matched the response of normal controls . For each experiment , we first determined which stimuli the patient recognized(see A . R. Oamasio et al., 1990 ). In brief , a recognition response was scored as correct if either of two conditions was met : 1 the stimulus was named correctly (we accept this as unequivocal evidence ( ) of correct recognition ; it should be noted that we have never found a subject who would produce a correct name and not recognize the stimulus that was named); or (2) the subject provided a specific description of the entity (e.g ., " ' that s an animal that can store water in the hump on its back, lives in the

276

Tests andClinicalSyndromes Neuropsychological

" desert , can go a long time without water, and can be ridden ). The number of stimuli the subjectrecognizedcorrectly, divided by the total number of items . score in the category and multiplied by 100, constituted the recognition The naming score was calculated by summing the number of correct a naming responsesusingonly thosestimuli for which the subjecthad produced . In other words, if a subject did not recognize a correctrecognition response was not included in the naming scorecalculation that stimulus stimulus , particular . In this approach to data quantification , subjectsare not penalizedfor . Scoreswere multi also do not that name stimuli to they recognize failing plied by 100 to produce final figures in terms of percent correct. This clearly the recognition capacity of approachto data quantification separates the subject from the naming capacity of the subject , and avoids mischaracterizinga subject with a naming defect as having a recognition defect. Scoresfrom the brain-damaged subjects were compared with those of the controls, and scoresthat fell 2 or more SO below the mean of the control group were classifiedas defective.
The neuroanatomical analysis was based on Methods Neuroanatomical magnetic resonance (MR ) data, or in those subjects in whom MR imaging could not be obtained , on computezed tomography (CT ) data, obtained at least three months post onset of lesion . In most subjects, MR scans were obtained with an SPg sequence of thin ( 1.5 mm ) and contiguous T I -weighted coronal cuts. The resulting 124 slices : Frank, 1992 ) and a three were processed in Brainvox (H . Damasio & dimensional (3- D ) reconstruction was obtained for each individual subject . In a few cases only a standard MR sequence was available with both axial and coronal T I -weighted , 5-mm -thick slices. The anatomical description of the lesion and of its placement relative to neuroanatomical landmarks was performed with Brainvox . The results of the analysis were stored in relation : brain segmentation used in our laboratory (H . Damasio & standard to the Damasio , 1989 ). In a subsequent step, all lesions in this set were transposed and warped into a normal 3-D brain , so as to permit the determination of the maximal overlap of lesions relative to subjects grouped by neuropsychological defect. This technique , known as MAP -3, is carried out as follows : The normal 3 -D brain is resliced so as to match the slices of the MR or CT of the subject and ' create a correspondence between each of the subject s MR or CT slices and the normal resliced brain . The contour of the lesion on each slice is then transposed onto the matched slices of the normal brain , warping it in relation to the available anatomical landmarks . The summation of these contours " " defines an object which represents, for each subject , the lesion in three dimensions placed in the normal reference brain . The final step consists of " " the detection of the intersection of the several objects in the reference brain , and the analysis of their placement in relation to the anatomical detail

277

: The Neural Basisof ~ xical Retrieval : Damasio Tranel, Damasio ,&

of the reference brain . If the analysis is only concerned with the surface overlap of the lesions , we have designated it as MAP - 2. In t~ s instance, each view of the 3- D brain showing the lesion is matched , for each subject , with the corresponding view of the normal brain . The contour of the lesion is ' transposed from the subject s brain onto the normal brain taking into account its relation to sulcal and gyrallandmarks .

ResultsConcerningHypothesisNo. 1 Retrieval of words vs. retrieval of for the same concrete entities . concepts
Neuropsychological Findings Table 12.1 presents neuropsychological results pertinent to the question of whether retrieval of conceptual knowledge and retrieval of words can be dissociated . So far, we have uncovered three different patterns of recognition and naming profiles, with respect to nonunique concrete entities: (1) group 1: defective naming accompaniedby normal recognition; (2) group 2: defective naming accompanied by defective have defective in and addition , , have defective recognition (subjects recognition ); (3) group 3: normal naming but naming of items they can recognize defective recognition (as defined for group 2), that is, subjects who could name correctly entities that they recognized , even though overall recognition was defective in one or more categories . Data &om the normal controls are also presented . NeuroRnRtomicRI Findings Abnonnal retrieval of words and nonnal retrieval of concepts (see table 12.1, 1 . The ten in this subjects group ) group had lesions in the left temporal lobe, mostly overlapping in the middle and lateral inferior temporal (IT) region, and in the lateral aspectof the occipitotemporal parietal region.

278

NeuropsychologicalTests and Clinical Syndromes

Abnonnal retrieval of words and concepts (see table 12.1, group 2). The twelve subjectsin this group showed lesions mainly in the left lateral temporo- occipital region, mostly overlapping at the temporo- occipitoparietal . junction, at the posterior end of the superior temporal sulcus Nonnal retrieval of words and abnonnal retrieval of concepts (see table 12.1, group 3). Of the twenty one subjectsin this group, eleven had lesions in the right hemisphere , and ten in the left. The right unilateral lesions were concentratedin the inferior and mesialaspectsof the occipital lobe. The main overlap of lesions was in the infracalcarineregion (lingual gyrus), tapering , there was anteroinferiorly in the posterior IT region. In the left hemisphere an overlap of subjectsin the anterior sector of the fusiform gyrus (seeT ranel, Damasio , 1997b). , & Damasio Results for hypothesis No . 2 Retrieval of words for concrete entities

from separate conceptual categories .

Rl Findings In table 12.2, the results are organized Neuropsychologic according to the category of naming impairment (collapsedacrosswhether or not there was an accompanyingrecognition defect ). In group 1, subjects -naming defect had an animal ; in , irrespective of other naming impairments animal category to the 1 defect was restricted the 2 a subset of group ), group ( . In group 3, subjectshad a tool and utensil naming defect , irrespective of other naming impairments ; in group 4 (a subset of group 3), the defect to the tool and utensil category. was restricted We conducted statistical comparisons to confirm the reliability of the , we comparedthe naming scoresof defective groups findings. Using t-tests with the naming scoresof brain- damagedsubjectswho were not defective. The results supported the conclusion that there were significant differences -naming score differed significantly between groups. For group 1, the animal

279

: The Neural Basisof Lexical Retrieval : Darnasio Tranel, Darnasio ,&

from the score of nondefective subjects (t ( 125) = - 4.79, P < .001). For group 3, the tool -utensil- naming score differed significantly &om the score of nondefectivesubjects(t ( 125) = - 4.56, P < .001).
Neuroanatomical Findings Abnormal retrieval of words for animals (see table 12.2, group 1). A defect in retrieval of words for animals was observed in sixteen subjects. In all but one, the lesions occurred in the left IT region . The maximal overlap was seen in the midanterior sector of the lateral and inferior aspect of the IT region . The overlap then tapered toward the anterior sector of the IT region and the temporal pole . With the exception of one subject , whose lesion was in the mesial left occipital region (in both supra- and infra calcarine regions ), no lesions outside the IT region were associated with the defect. Abnonnal retrieval of words for tools and utensils (see table 12.2, group 3 ). Abnormal retrieval of words for tools and utensils was associated with damage in the left lateral temporal and occipital region and, to a lesser extent , in the left parietal region . (One subject had a lesion in the right temporal pole .) The maximal lesion overlap occurred in the left temporo - occipitoparietal junction .

Fundional Imaging (PET ) Studies


We recently completed a functional imaging study regarding retrieval of words for concreteuniqueand nonuniqueentities (A . R. Damasio , Grabowski, Damasio , et al., 1995; H. Damasio et al., 1996). We studied nine normal right -handedyoung adults, ranging in age from 22 to 49, all of whom were native English speakers . There were seven women and two men. The subjects : (1) naming unique personsfrom their faces engagedin three tasks ; (2) animals and 3 tools and utensils ; results for the ( ) ( naming naming naming of familiar faces are below . In a control ) task , subjects were unique reported asked to decide and report whether unfamiliar faces were presented right side up or upsidedown. The naming taskswere performed during a PET scanningsession . Subjects each task twice in random order. Task , performed performancebegan 5 seconds after injection of [lSO]H1O into the antecubitalvein and continued until 65 secondsafter injection. Oral responseswere recorded , and performance measures(accuracy , latency) were obtained. For each task, the stimuli were presentedat set rateswhich pilot studieshad shown to yield similar high but , that is, rates at which subjectscould perform nonperfect performanceaccuracies well, but not at ceiling. Specifically , the familiar faces were presented one every 2.5 seconds , the tools and utensils were presentedone every 1.8 seconds , and the animalswere presentedone every 1.5 seconds . Theserates levels of about 90 percent correct for each task type. produced performance ' In a separatesession , MR imagesof each subjects brain were obtained , and

280

aJ Tests andClinical Neuropsychologic Syndromes

1 : Frank reconstructed in three dimensions using Brainvox (H. Damasio & , 1992). -Brainvox (GraMR and PET data were coregistered a priori using PET bowski, Damasio , Frank et al., 1995). This fit was corrected post hoc with 1 : Cherry 1993). PET automatedimage registration (AIR ; Woods, Mazziotta, & 1 : Szikla data were subjectedto Talairach transformation (Talairach & , 1967), . The data were analyzed based on analysis of the coregistered 3-D MR dataset with a pixelwise two -way, analysis of covariance (ANCOV A ; estimated coefficients for global flow serving as the covariate), in which we comparedadjustedmeanactivity in eachof the three naming conditions with the control task (Friston, Fritch, Liddle, et al., 1991). Regions of statistically significant changes in normalized regional cerebral blood flow (rCBF) for each of the three naming tasks were searchedfor in the temporal polar (TP) area and in the IT cortices identified in the 3-D reconstructedMR scansof eachsubject . When subjectsnamed animals and tools and utensils , there were significant increasesin rCBF in distinct loci in the posterior IT region in the left . For the tools and utensils , the principal location of activation hemisphere was in the posterolateralaspectof the left IT region, in the middle and inferior temporal gyri . Naming of animalsproduced activation which was mesial and anterior to that produced by naming tools and utensils , in the inferior and fourth temporal gyri . Most important, the areasof activation produced , but also by naming animals and tools and utensils were not just separate studies as the lesion areas identified to the by being crucial to corresponded . Our results are consistent with another recent PET study these capacities (Martin et al., 1995).
Conclusions

The data summarizedabove indicate that the neural systems required to retrieve conceptualknowledge for nonunique entities, and those required to retrieve the words for those entities, are separate , at least in part. The separation is most noteworthy for the animal category. For the tool and utensil ; in fact, basedon findings available category, the separationis less apparent . The results also thus far, it is not really possible to separatethe systems address the status of the fruits and vegetables category. In nearly all of our subjects , the outcome for this category paralleled the outcome for the animals category. For example , in subjects with severely impaired animal fruits and vegetableswas also severely defective of , recognition recognition in most cases(seegroup 3 in table 12.1). In fact, only two subjectsout of a total of thirty -three (groups 2 and 3, table 12.1) had defective animal recognition but normal fruit and vegetablerecognition. Also , in subjectswho had , the magnitude of impairment was naming defects in all three categories and and fruits in the animals vegetablescategories(groups 1 and comparable 12.2 . 3, table )

281

: The Neural Basisof Lexical Retrieval Tfane), Damasio I: Darnasio ,&

The data support the hypothesis that the neural systems required to retrieve words for nonunique entities are based nearly exclusively in left hemisphereregions. We only found one subject (out of 127) in whom defective word retrieval occurred with a lesion in the right hemisphere . The data also support the hypothesis that there are separableneural regions special ized for word retrieval related to entities from different conceptual categories . Abnormal retrieval of words for animals occurred in subjects with lesions that clustered in the anterior portion of the IT region. The maximal overlap occurred in lateral and inferior IT regions, in the anterior sector of the middle and inferior temporal gyri . Of importance with regard to the , the temporal polar region was not findings discussedin the next section included in the overlap. Abnormal retrieval of words for tools and utensils occurred in subjects whose lesions involved the posterior and lateral temporal cortices and the supramarginalgyrus; the maximal overlap occurred at the back end of the middle temporal gyrus and the anteroinferior sector of the supramarginalgyrus (a region we have designatedas " posterior IT + " ). It is interesting to note that the system segregation effects noted in our studiesseemto obey consistentprinciples. Specifically , relative to the stream architectureof cortical projection neurons in the occipitotemporal region, (a) word retrieval defects for animals depend on damage to systems located anterior to those whose damage produces tool and utensil word -retrieval defects ; and (b) in no instance was there a defect for word retrieval for animalscausedby a lesion posterior to a lesion causingword -retrieval defects for tools and utensils . In other studies , we have found category-related neuropsychologicaland neuroanatomicaldissociationsregarding the retrieval of conceptual knowledge for concrete entities, which parallel to some extent the dissociations described here for word retrieval (A . R. Damasio , 1990; A . R. Damasio et al., 1990; Tranel, Damasio , Damasio : Damasio , et al., 1995; Tranel, Damasio ,& , 1997b). For example , in a large-scalestudy of subjectswith lesions throughout various sectorsof the telencephalon , we found that damagecenteredin the right mesial and inferior occipitotemporal region, or centeredin the left mesial occipital region, produced impairments in the retrieval of conceptual . Damage centered in the left posterior knowledge (recognition) for animals and temporo occipital parieto occipital regions produced impairments in the retrieval of conceptualknowledge for tools and utensils . Thesefindings suggest that, as in the caseof words, the recording and retrieval of conceptual knowledge of different domains depend on partially segregatedneural systems : Damasio , Damasio ( Tranel ,& , 1997b ).

RETRIEVAL OF WORDSFOR UNIQUE CONCRETE E niTfi ES Background The retrieval of words for unique , concreteentities(e.g., persons , places ) relative to the retrieval of words may be disproportionately compromised

282

a J Tests andClinical Neuropsychologic Syndromes

for nonunique entities, and may in some casesconstitute the only wordretrieval defect (Carney & Temple, 1993; Cipolotti , McNeil , & Warrington , 1993; Cohen , Bolgert, Timsit, et al., 1994; A . R. Damasio , 1990; Flude , Ellis, & Kay, 1989; Hittmair -Delazer , Denes , Semenza , et al., 1994; Lucchelli & De Renzi , 1992; McKenna & Warrington, 1978, 1980; McNeil , Cipolotti , & , Warrington, 1994; Semenza& Zettin , 1988, 1989; Shallice & Kartsounis 1993). However, most of these studies were conducted with little or no regard for neuroanatomicalfactors. This has produced a rather confusing array of findings, as far as understanding the neural basis of retrieval of words for unique entities is concerned . In fact, Hittmair -Delazer et al. (1994) concludedafter reviewing this literature that it was not possible to provide a " retrieval. The neuroanatomicalbasis for " proper name authors provided a " table showing the various lesion locations for the casesof " proper anomia " that have been reported to date, and concluded that the relative pureness " " of all reported cases and the observation of quite distinctly located ana " tomical lesions made it unlikely that an anatomical explanation would suffice . However, we believe this interpretation is a consequence of either insufficient precision of neuroanatomicalanalysis or inadequateassessment of naming compromise , or perhapsboth. In fact, we have findings obtained with careful neuroanatomical and , which point neuropsychologicalinvestigations in a large series of patients and a to neural correlate for retrieval of words consistently strongly specific for unique entities (A . R. Damasioet al., 1990; H. Damasioet al., 1996; GraffRadford et al., 1990). The findings also suggest that accessto words for unique entities dependson neural systemsdistinct from those that support access to words for nonunique entities (A .R. Damasio & Damasio , 1992; A . R. Damasioet al., 1991; A . R. Damasioet al., 1995; H. Damasioet al., 1996). Finally, the findings indicate that, as in the case of nonunique entities, the neural systems supporting retrieval of conceptual knowledge are partially from those supporting the retrieval of words { Tranel , Damasio ,& segregated Damasio , 1997b ).

New Lesion Studies


The main hypothesis we addressedin these experiments is that accessto words for unique entities dependson neural systemsdistinct from the neural to words for nonunique entities. systemsthat support access
We studied subjects with damage to varied sectors of the left or Subleds right hemisphere , as described in the previous section (n = 127 ). Fifty -five normal controls , described in the previous section , were also studied .

Method Stimuli from two tests were utilized to measurenaming of famous faces : seventy-seven items from the Iowa FamousFacesTest (Tranel, Dama sio, & : Damasio , 1995), and fifty -six items modified from the Boston Famous

283

Tranel, Damasio : Damasio : The Neural Basisof Lexical Retrieval ,&

Table 12.3 Retrieval of words for unique entities -

Group BrRin - damagtd subleds Normal Defective (all) Defective (persons only) Normal controls

N 114 13 7 SS

Mean 89.3 49.6 63.4 92.3

(SO ) (4.9) (23.1) .0) (10 (6.2)

FacesTest (Albert , Butters : Levin, 1979). The stimuli are black-and-white ,& slides depicting famous actors . The slides , politicians, and sports persons were shown one at a time, and for eachthe subject was askedto (1) indicate whether the face is familiar, and if so, to (2) indicate who the person is, and ' . (3) supply the persons name
Data Quanti6cation The data were quantified in the same manner as described above for nonunique entities . Specifically , for each subject we first determined the stimuli for which the subject produced unequivocal recognition responses (using the same criteria as specified in the previous section ). For example , for the pertinent stimulus , "Michael Jordan" and " famous basketball player who tried to play baseball; his team won more games than any other team during the season" would qualify as correct recognition " " " responses, whereas sports guy and basketball player " would not . Then , for the stimuli that were recognizedcorrectly, we determined the naming score by dividing the number of correct naming responses by the number of correct recognition responses. The naming score was multiplied by 100 and compared with scores from controls ; scores 2 or more SO below the control mean were considered defective .

Neuropsychological Results The results are summarizedin table 12.3. There were thirteen subjectswho had defective retrieval of namesfor persons . The mean performancein this group was significantly lower than that of the brain- damaged subjects who were not defective (t ( 125) = - 6.17, P < .001). Of the thirteen defective subjects , seven had a defect that was ' exclusive for the persons category, that is, they had defective retrieval of ' s but words for person normal retrieval of words for animals , , fruits and . Also, it is important to reiterate that none , and tools and utensils vegetables of the thirteen subjects with defective retrieval of words for persons had defective recognition of persons . In fact, the averagerecognition performance in the thirteen subjectswas 81.9 (SO= 9.8), which comparesfavorably with the control averageof 75.7 (SO = 6.7).
Neuroanatomical Results The techniQues for analvsis were the same a those reported for the neuroanatomical results in the study above regarding
~ ~

284

NeuropsychologicalTests and Clinical Syndromes

retrieval of words for nonunique entities . In the seven subjects with a pure defect in retrieval of words for unique entities , damage was centered in the left TP region , in both the inferomesial and lateral aspects. When all subjects with a deficit in retrieval of words for persons were considered (n = 13), the site of maximal lesion overlap remained in the anterolateral and inferior sectors of the left TP region . Functional

The methods for the functional imaging study of proper naming were described in the previous section . When the nine normal subjectsnamedfamiliar faces , there was an increasein normalized rCBF in the left TP but not in the left IT region. The left TP activation correspondsto the sameregion identiBed in the lesion study as being important for the retrieval of words for , which we unique entities. There was also activation in the right TP area would to the of the faces that reflect inevitably recognition unique interpret the control did not the of the faces task ( naming require recognition accompany at a unique level). Again, this Anding closely parallels results from with naming unique entities lesion studies . The left TP activation associated was distinctfrom the activation in IT associatedwith naming of nonunique entities. This Anding provides convergent evidencefor the existenceof distinct anatomicalsystems supporting retrieval of words for unique vs. nonunique entities.
Conclusions Our findings indicate that lesions to the left TP region were associated with defects in retrieval of words for unique concrete entities . None of these subjects , however , had a defect in retrieving the concepts of the persons for whom they could not retrieve the names, that is, recognition of identity was normal . We also found that lesions restricted to the left IT region , which did not involve the temporal pole , did not cause defects in retrieval of words for unique entities . Rather , as noted in the previous section , these subjects had impairments in word retrieval for nonunique entities . Weare intrigued that the findings continue to obey the principles described earlier in connection with our studies regarding word retrieval for nonunique entities . Specifically , word retrieval for unique entities depends on a system located anterior to the systems on which word retrieval for nonunique concrete entities depends. Of all the subjects with naming impairments (n = 30 ), there was never a violation of this principle . For example , there was never a case in which a naming defect for unique entities related to a lesion located posterior to a lesion that caused a defect in the naming of nonunique entities . Also , there was never a combination of a defect in naming unique entities with a defect in naming nonunique entities of the tool and utensil variety , without a defect in naming animals as well . In other

285

: TheNeuralBasis of Lexical Retrieval Tranel : Damasio , Darnasio ,&

words, we never found that a subject could have a defect in naming persons and in naming tools and utensils . , but could be normal in naming animals In sum , impaired retrieval of words for concrete entities correlated with . Impairments damagein higher-order cortices outside classiclanguageareas in various word categorieswere correlated in a consistent manner with separable neural sites , words for persons with TP lesions , words specifically for animals with anterior IT lesions and words for tools and utensils with , IT lesions . It is notable that two of these regions TP and posterior posterior + IT + - are not contiguous , do not overlap cortically or subcortically , and are so distant as to make it virtually impossible for a single lesion to compromise them without also compromising the intervening region. This explains why a combined defect for personsand for tools and utensils never occurredin our sample .

RETRIEV At OFWORDS FORACTIONS


Background

We have recently begun to study the neural correlatesof retrieval of conceptual . In one study (A . R. Darnasio knowledge and words regarding actions & : Tranel, 1993), we reported a subject with a well-definedlesion in the left frontal operculum , involving both prefrontal and premotor cortex and white matter. The subject had defective retrieval of words for underlying actions but normal retrieval of words for nonunique entities (animals , , tools and utensils ). This subject was contrasted with two other subjects , who had left IT lesions . Those two subjectshad the reverse pattern- defective retrieval of words for concrete entities but normal retrieval of words for actions . Regarding the first subject and the other two , the lesions were nonoverlapping and the neuropsychological performanceswere quite distinct, and the result thus constituted a double dissociationrelative to both naming performanceand site of lesion. This result hinted that the systems required for naming of entities and naming of actions are segregatedin the human brain, at least in part, even if they normally operate in coordinated fashion. Our result came on a background of observations that patients can have disproportionate impairment " or " verbs" of the ability to retrieve " nouns : Martin , 1996; (Breedin & Caramazza & : Hillis , 1991; De Renzi & : Pellegrino : War, 1995; McCarthy & : Caramazza ,& , 1994; Miceli , SilveriNoncen rington , 1985; Miceli , Capasso tini , et al., 1988; Miceli , Silveri , Villa , et al., 1984; Zingeser & : Berndt, 1988, 1990). Two recent reports (Daniele Givstolisi i Silver et al. , , , , 1994; Miozzo , Soardi & : , , 1994) support the Damasio and Tranel (1993) study in Cappa three and one patients . , respectively Another recent study compared syntactic constructions between patients ' with agrammatic(usually Broca s) aphasiaand those with conduction aphasia

286

NeuropsychologicalTests and Clinical Syndromes

: Gallagher , Christiansen , & , 1994). Consistent with previous (Goodglass studies of this type (e.g., Marin , Saffran : Schwartz , & , 1976), the authors found that patients with agrammaticaphasiahad a preponderanceof nouns over verbs inrunning speechand in single-constituent utterances . This finding is consistent with our work (e.g ., Darnasio & : Tranel, 1993), and with other studies reporting noun-verb discrepancies in word repetition (Katz & : : Warrington , 1985; , 1990) and written word retrieval (Baxter & Goodglass Cararnazza & : Hillis , 1991). We have replicated and extended our initial finding in several additional cases , as summarizedbelow.
New Lesion Studies We have studied eighty -three subjects with lesions in the left or premotor prefrontal region , left temporal , parietal , or occipital cortices , or various right hemisphere regions . The subjects were drawn from our patient registry , and they conformed to the same inclusion criteria specified for the studies on concrete entities described earlier . Subleds

Experimental Task The task for measuringretrieval of words for actions was the Action Recognition and Naming Test recently developed in our : Tranel, 1997). In brief, the test comprises 280 color laboratory (Fiez & . The items elicit responses of various actions which vary along photographs severaldimensions , including (1) the inflection of the elicited response(gerundial forms [e.g., eating ] vs. past-tense forms [e.g ., ate ]), (2) the &equencyof the elicited verb per million words (Francis& : Kucera , 1982), (3) the type of , animal , or object), and (4) compatibility agent performing each action (person with different argument structures can ; that is, some elicted responses only be produced in a well-formed sentenceas an intransitive verb (one: e.g., "John RAN /IS RUNNING " ), some only as a transitive place predicate " " : e.g ., John HIT /IS HITTING the ball ; or threeverb (two -place predicate " : e.g., John GAVE Mary a book), and some can be produced place predicate in either type of sentence . The items also representa diverse range of conceptual e. . verbs of perception , , motion, etc.). In the test, 75 percent ( categories g of the stimuli are single pictures depicting an ongoing action, for which subjectsare instructed to produce a single word which best describeswhat " " the person , animal , or object is doing (e.g., walking ). The remaining 25 percent of the stimuli are picture pairs depicting some change in an object, and subjectsare asked to produce a single word which best describeswhat ". was done to the object, or what the person or object did (e.g ., " chopped ) Scoring and Data Quantification The Action Recognition and Naming Test was standardizedin a series of experimentsconducted in normal subjects : Tranel, 1997). This yielded, for eachitem on the test, a naming (Fiez&

281

Tranel, Darnasio : Damasio : The Neural Basisof Lexical Retrieval ,&

Table 12.4 Summary of actionrecognition andnaming data(no. of subjects defedive ) Lesion group Leftfrontal Leftocdpitotemporal Rightocdpitotemporal Other Total TotalNo. 19 21 9 34 83 (and Naming ) Recognition Naming Only

10 8 1 2 21

0 2 1 1 4

, two or three responses ) which is considered response(or in a few instances correct. To quantify the performancesof the brain- damagedsubjects , we , and compared their responseswith those of the standardization sample scored as correct those responses that matched those produced by the normal controls. A percent correct score was then calculated for each brainif it was 2 or more , and the score was classifiedas defective damagedsubject SO below the meanof the control group. A hierarchical approach was used to classify the brain-damaged subjects into four different groups. First, we identified a group of nineteen subjects whose lesions included (but were not necessarilylimited to) damageto any of the following left frontal areas : left frontal operculum , premotor region, rolandic region, basal ganglia. From the remaining sixty -four subjects , we next identified a group of twenty -one subjectswhose lesions included (but were not necessarilylimited to ) damage to any of the following left occi: infracalcarine cortex, supracalcarinecortex, the mesial pitotemporal areas , the posterior portion of the middle temporal junction temporo occipital the , gyrus posterior portion of the IT gyrus. From the remaining forty -three we next identified a group of nine subjectswhose lesions included , subjects (but were not necessarilylimited to ) damageto the mesial occipitotemporal areas in the right hemisphere . The remaining thirty -four subjects (whose lesions did not extend into any of the areaslisted above) were not classified into specificneuroanatomical . subgroups Results The results are presented in table 12.4. We found a number of subjectswho had a defect in the retrieval of words for actions. It is interesting to note that in most cases , this impairment was accompanied by a defect in the retrieval of conceptualknowledge for actions (the " naming and recognition " group). Thesefindings provide a replication in ten addiHonal subjectsof our initial result regarding the associa Hon of defective retrieval of words for actions with damage to the left premotor or prefrontal region (Damasio & : Tranel, 1993). Neuroanatomicalanalysis in these subjectsrevealed that the lesions clustered in the inferior motor and premotor regions, with the maximal lesion overlap occurring in the inferior frontal gyrus and the inferior sector

288

Tests andClinic a JSyndromes Neuropsychological

of the precentral gyrus. The overlap tapered both into more anterior prefrontal regions, and posteriorly into the supramarginalgyrus. The findings also suggest a strong associationbetween recognition and naming, with regard to actions. In the brain-damagedsubjectswith defective naming, we found that there was a significant correlation between naming performanceand the averageconceptualretrieval (recognition) performance (r = .55, P < .005). Thus, recognition and naming tend to go together quite strongly with regard to actions. This contrasts with the frequent dissociations we have obtained betweenthesetwo capacitieswith regard to concrete entities, especiallypersonsand animals(although it is interesting to note that the tools and utensils category is more like the actions , inasmuchas naming and recognition tend to be disturbed concomitantly).

) Studies Fundionallmaging (PET


" There is a PET paradigm known as verb generate ," in which subjects are Petersen Fox verbs action words to , , Posner , et al., 1988, ) ( ( required generate 1989; Raichle , Videen, et al., 1994). We have conducted a PET study , Fiez which replicates and extends the Andings from the verb-generate studies , Frank , Brown, et al., 1996). We studied eighteen normal righthanded (Grabowski volunteers , who underwent both a 3-D MR study and a PET study, using the samemethods describedearlier. The subjectsreceivedinjections of , subjectswere [lSO]H2O while performing the verb-generatetask. In the task , and presentedcommon nouns visually at the rate of one word per 2 seconds for each ) a verb that went with the , the subjectshad to generate(and speak noun. Changesin rCBF associatedwith the task were analyzed using the samemethodsdescribedearlier. The strongest areas of rCBF increaseassociatedwith the verb-generate task were the left inferior frontal gyrus, left dorsolateral prefrontal cortex, . These regions correspondclosely to those reported in and right cerebellum other PET studies using the same paradigm (Petersen et al., 1988, 1989; Raichleet al., 1994; see also Martin et al., 1995). Also, the Anding of rCBF in the left premotor and prefrontal region during the PET studiesis increases . Convergent evidence quite congruent with the Andings from the lesion studies is also available from functional studies using other imaging modalities, suchas functional MR imaging (Hinke, Stillman, Kim, et al., 1993).
Conclusions

The lesion studies support the idea that defective retrieval of words for actions is correlatedwith damagein the dorsolateral sector of the left frontal . lobe, mostly in the inferior frontal gyrus, in premotor and prefrontal cortices Functional imaging studies have produced findings consistent with this . conclusion

289

: The Neural Basisof Lexical Retrieval Tranel, Darnasio : Oarnasio ,&

CONCLUDING REMARKS
The findings summarizedin this chapter indicate that the retrieval of words which denote concrete entities dependson regions in higher-order association cortices in the left temporal region, and that the retrieval of words denoting actions depends on the left prefrontal and premotor region. We have suggestedthat these regions play an intermediary or mediational role in lexical retrieval (H. Damasioet al., 1996). For example , when the concept of a given animal is evoked (basedon the activation of regions which support pertinent conceptualknowledge and promote its explicit representation in sensorimotorterms ), an intermediary region becomesactive and promotes (in the appropriate sensorimotor structures ) the explicit representation of to the word form which denotes the given phonemic knowledge pertaining animal. A different intermediary region is engaged when a concept from - is evoked. This another category- a tool , for example processcan operate in reverse to link word form information to conceptual knowledge. We do not believe that the intermediary regions contain in explicit form the names for all entities; rather, we suggest that they hold knowledge about how to reconstructa certain pattern (e.g ., the phonemic structure of ;-. given word ) in . explicit form, within the appropriate sensorimotor structures We have hypothesized that there are two main reasonswhy naming of different kinds of entities is correlated with different neural sites . One pertains to the overall physical characteristics of the entity being named , which determine the kind of sensorimotor mapping generated during interactions between an organism and the entity , and which are a key to the neural mapping of the correspondingconceptualknowledge. The other reasonpertains to the fine physical characteristics and contextual linkagesof an entity which . We permit the mapping of unique items such as familiar persons or places presume that the varied conceptual specificationwhich results from factors such as physical characteristicsand contextual complexity is the driving force and " principle" behind the differential and consistent neural placement of lexical intermediary regions. For instance , the multiple sensory channels (somatosensory , visual) and the hand motor patterns which are inherent in the conceptualdescription of a manipulabletool would be a driving force for the anatomical gravitation of the respective intermediary region toward a sector of cortex that is capableof receiving such multiple sensory signals , and is close to regions involved in visual motion and hand motion processing . The driving force for the intermediary region concerning names for personscomprisesboth the finer physical specificationand cornextual complexity to define suchunique items. necessary As far as the possible microstructure of the intermediary regions, we do " or " not envision them as rigid " modules hard-edged " centers , becausewe seetheir structure and operation as being acquiredand modified by learning. We suggestthat an individual' s learning experienceof conceptsof a similar kind (e.g ., concepts of manipulable tools), and their corresponding words,

290

NeuropsychologicalTests and Clinical Syndromes

leads to the recruitment , within the availableneural architecture , of a critical set of spatially proximate microcircuits. We presume that the anatomical placement of the region within which the microcircuits are recruited for a certain range of items is not haphazard . Rather , it is the one best suited to permit the most effective interaction between the regions of cerebralcortex that subtend conceptualknowledge and those that can enact lexical knowledge . The findings in our studies suggest that distinct kinds of conceptual knowledge lead to the recruitment of distinct intermediary regions. We expect normal personsto develop, under similar conditions, a similar type of , but we also predict ample individual variation of large-scale architecture within each microcircuitry key region. We also expect that, at different times, the sameindividual will engagethe samelarge-scaleregions, but not neces sarily the samemicrocircuitry within them. In sum , the findings describedhere offer support for the idea that, inaddition to the separationof the neural systemswhich support the retrieval of , there is a parallel concepts for entities belonging to varied categories regionalization for the systems which support the retrieval of the word forms correspondingto those entities.

ACKNOWLEDGMENT
Supported by NINDS Program Project Grant NS 19632 and ONR Grant -91-J- 1240. The authors thank Joan Brandt and Jon Spradling for NOOO14 technicalhelp with the experiments , and Denise Krutzfeldt for help in scheduling the subjects .

REFERENCES
Albert ., Butters. N., & : Levin in the retrograde , M. S amnesia , J. A. (1919 of ). Temporal gradients
' . Archivesof Neurology , 36, 211- 216. patients with alcoholic Korsakoff s disease Basso ni, E., & : Laiacona , A ., Capita , M . (1988). Progressivelanguageimpainnent without dementia : A casewith isolated category-specificsemanticdefect. ] oumRlof Neurology , Neurosurgery , and

- 1207 . , 51 , 1201 Psychiatry

-speci BaxterD. M., & : Warrington . (1985 Acphonological . Neuro , E. K ). Category dysgraphia - 666. , 23 , 653 psychologia Breedin : Martin of verb impainnent in aphasia : An analysis , S. D., & of , R. C. (1996 ). Patterns - 91. four cases . Cognitive , 13 , 51 Neuropsychology : Hillis, A. (1991 of nouns andverbsin the brain . Nature Carammaza , A., & ). Lexical , organization 349 - 790 . , 788 -speci : Temple A possible , A., & , C. M. (1993 ). Prosopanomia1 Carney category Acanomiafor faces . Cognitive 10 185 195 . , , Neuropsychology ., & : Warrington . (1993 writtennaming of propernames , L., McNeil : , J. E , E. K ). Spared Cipolotti A case . 1 331 . 289 , , Memory report Cohen .. Bolgen ., & : Chennann afterleft ,L , F., Timsit ,S , J. F. (1994 ). Anomiafor propernames - 1284 thalamic infarct . Journal . , Neurosurgery , andPsychiatry , 57 , 1283 of Neurology

291

Tranel, Darnasio : Oamasio : The Neural Basisof Lexical Retrieval ,&

-levelproposal -locked retroactivation : A systems for Damasio . (1989a ,A R ). Time multiregional of recall andrecognition . Cognition - 62. theneural substrates , 33 , 25 in thebrain . Mind andLAngUAge Damasio , 4, 24- 28. , A. R. (1989b ). Concepts -related asa clueto theneural substrates of defects Damasio , A. R. (1990 ). Category recognition in Neurosciences . Trends , 13 , 95- 98. knowledge - 539 . NewEngland . Damasio , 326 , 531 , A. R. (1992 ). Aphasia ] oumal of Medicine : Retrieval . , Brandt : Damasio of Damasio , J. P., Tranel , D., & , H. (1991 ,A R ). Namedropping in left cortex . or common nouns on different for Neuro systems Society depends temporal proper sama Abstracts , 17 , 4. . Scientific American Damasio : Damasio , 267(September , A R., & , H. (1992 ), ). Brainandlanguage 88- 95. : The for retrieval of concrete Damasio : Damasio , H. (1994 , A R., & ). Cortical systems knowledge -scale theories . In C. Koch(Ed .), LArge neuronal zoneframework of thebrain(pp. 61convergence : MIT Press . 74 , MA ). Cambridge of Damasio . , Damasio : Brandt , D., & , J.P. (1990 , A. R , H., Tranel ). The neuralregionalization . evidence . Quantitative - 1047 access : Preliminary , 1039 , 55 Biology knowledge : Anatomic basis and : VanHoesen Damasio , A R., Damasio , H., & , G. W. (1982 ). Prosopagnosia - 341 . . Neurology behavioral mechanisms , 32 , 331 Damasio , L. L. , H., Tranel , D., Frank , R. J., Spradling , A. R., Grabowski , T. J., Damasio , J., Ponto fromseparate lexical areretrieved B., Watkins : Hichwa , G. L , & , R. D. (1995 ). Separate categories . Society . : A PETactivation , 21 , 1498 study for Neuroscience systems " " " " Damasio : Tranel , A R., & , D. (1990 ). Knowingthat Colorado goeswith Denver doesnot " " " . " is in Colorado . Bthavioural BrainResearch that Denver , 40 , 193 200 implyknowledge with differentlydistributed Damasio ., & : Tranel , D. (1993 ,A R ). Nounsandverbsareretrieved - 4960 . . Proceedings neural , 90 , 4957 Academy of Sciences of theNational systems in neuropsychology . New York: Oxford Damasio : Damasio , A. R. (1989 , H., & ). Lesion analysis Press . University in humans . . (1992 -dimensional in vivo mapping of brainlesions Damasio : Frank ,R , H., & ). Three - 143 . Archives , 49 , 137 of Neurology : Damasio Damasio , A. R. (1996 , D., Frank , R. J., Hichwa , R. D., & ). A , H., Grabowski , T. J., Tranel - 505 . for lexicalretrieval . Nature neural basis , 499 , 380 for a i , M. C., Colosimo : Gainotti Daniele , C., & , G. (1994 , L., Silver ). Evidence , A , Giustolisi of nounsandverbs . Neuropsychologia basis for lexicalprocessing neuroanatomical , 32 , possible - 1341 1325 . -spedficword processing in the evidence for category Dehaene . (1995 ,S ). Electrophysiological - 2157 . NeuroReport . normal human brain , 6, 2153 in a of verbsandpreserved De Renzi : Pellegrino , but ineffectual , E., & , G. (1995 ). Sparing reading . Cortex - 636 . with impaired word production , 31 , 619 patient is not : McMullen , P. A. (1996 Farah , M. J., Meyer , M. M., & ). The living/nonlivingdissociation . Cognitive a strongtest anartifact : Givingana priori implausible , 13 , Neuropsychology hypothesis - 154 . 137 -bounded anomia : Implications for the neural : Wallace Farah , M. J., & , M. A. (1992 ). Semantically . 30 609 621. of , , Neuropsychologia naming implementation

292

Tests andClinicalSyndromes Neuropsycholo ~icaJ

: TraneLD. (1991 stimuli and procedures Fiez , J. A , & for investigating the ). Standardized - 569 retrieval of lexical andconceptual for actions . Memory andCogni Hon . , 25 , 543 knowledge Flude : Kay , B. M., EUis , A. W., & and nameretrievalin an anomic , J. (1989 ). Face processing : Names arestored from semantic information aboutfamiliarpeople . Brain Rnd aphasic separately . , 11 , 60- 12 Cognition Francis : Kucera Rn US : lit.ricon , W. M., & andgrRmm Rr. , H. (1982 ). Frequency Rlysis of English Rge Boston : Houghton Mifflin. FranklinS., HowardD ., - 566 . , 12 , 549 & : Patterson . Cogni HveNeuropsychology , K. (1995 ). Abstractword anomia

Friston . J., Frith : Frackowiak .S.}. (1991 ,K , C. D., Liddle functional , P. F., & .R ). Comparing (PET ) : The assessment of significant . ] ounlRl R I Blood FlowandMetRbolism , 11 , images of Cerebr change 690 - 699 . : Budin andmodalityspecific , H., & dissociations in word comprehension , C. (1988 ). Category Goodglass andconcurrent . 26 61 18 . , , dyslexia R phonological Neuropsychologi : Gallagher , H., Christiansen constructions usedby , J. A , & , R. E. (1994 ). Syntactic Goodglass : Comparison with conduction and normals . Neuropsychology , 8, agrammatic speakers aphasics - 613 . 598 : Theurkauf , H., Wingfield , A , Hyde dissociations , M. R, & , J. C. (1986 ). Categoryspecific Goodglass in naming andrecognition . Coria, 22 . , 81- 102 by aphasic patients -Ponto Grabowski , T. J., Damasio : Watkins , H., Frank , R, Hichwa , R. D., Boles , L L, & , G. L and MRI-PETcoregistration . Human Brain (1995 ). A new tedmiquefor PETsliceorientation - 133 . , 2, 123 Mapping -Ponto Grabowski . , Damasio , T. J., Frank : , R J., Brown , C. K , H., Boles , L L , Watkins , G. L , & R D. . 1996 of PET activation across statistical methods Hichwa , ( ) Reliability , subject , and groups . Human sizes Brain , 4, 23- 46. Mapping sample -Radford Graff , N. R , Darnasio , A R., Hyman , B. T., Hart , M.N., Tranel , D., Darnasio , H., Van 's disease Hoesen : Rezai . (1990 in a patientwith Pick : A neuropsychological , G.W., & ,K ). Progressive aphasia . Neurology . - 626 , radiologic , andanatomic , 40 , 620 study -speafic Hart : Cararnazza deficitfollowingcerebral , J., Berndt , R. S., & , A. (1985 ). Category naming infarction . Nature 316 . 439 440 , , Hart : Gordon for objectknowledge . Nature , J., & , B. (1992 ). Neuralsubsystems , 359 , 60- 64. -specific Hillis, A E., & : Cararnazza andcomprehension : , A. (1991 ). Category naming impairment - 2094 A double dissociation . Brain . , 114 , 2081 Hinke .-G., Merkle : Ugurbi , R. M., Hu, X., StiUman , A E., Kim , S , H., Salmi , R, & L K. (1993 ). 's area Functional resonance . Neuro , 4, magnetic imagingof Broca during internalspeech Rtporl 615 - 618. 's -DeIazer Hittmair : Mantovan , M., Denes , G., Semenza , c ., & , M.C. (1994 ). Anomiafor people names . Neuropsychologia . 32 465 476 , , . B., & : Goodglass : Analysisof a rare type of repetition Katz , R , H. (1990 ). Deepdysphasia - 185 disorder . Brain andLanguage . , 39 , 153 andbrain . Cambridge . , MA: MIT Press ). Image Kosslyn , S. M. (1994 Lesser : Uematsu . (1994 stimulation and language , R., Gordon . Journal , B., & ,S ). Electrical of - 204 Clinical . , 11 , 191 Neurophysiology Levelt intention toarticulation . Cambridge . , W. J. M. (1989 ). Speaking , MA: MIT Press : From

293

Tranel, Darnasio : The Neural Basisof Lexical Retrieval : Darnasio ,&

Levelt words in speech : Stages es and representations , W. J. M. (1992 ). Accessing , process production . Cognition , 42 , 1- 22. - 230 Lucchelli : De Renzi name anomia . Coria, 28 . , F., & , E. (1992 ). Proper , 221 . M., Saffran : Schwartz of language in aphasia : MarinD . S , E. M., & , M. E. (1976 ). Dissociations - 889 for nonnalfunction . Annals . , 280 , 868 of theNewYorkAcademy of Sciences Implications Martin . V., Lalonde : Vngerleider cortical , A , Haxby ,J , F.M., Wiggs , C. L., & , L G. (1995 ). Discrete associated with of color and of . action Science 270 102 105 . , , regions knowledge knowledge Martin : Haxby of category , A., Wiggs , C. L , Vngerleider , L, & , J. V. (1996 ). Neuralcorrelates . NRturt - 652. , 379 , 649 knowledge specific , 8. M., Tzourio , N., Frak , V., Syrota , A , Murayama , N., Levrier , 0 ., Salamon , G., Mazoyer ., Cohen Dehaene : Mehler of speech . Journal ,S , L, & , J. (1993 ). The corticalrepresentation of - 479 Neuroscitna . , 5, 467 Cognitive . A., & : Warrington in anagramrnatic : The ,R , E. K. (1985 ). Category McCarthy specificity patient relative of verb retrieval and . 23 709 727 . , , Neur Dps Ychologia impainnent comprehension -specific . A,& : Warrington for modality in ,R , E. K. (1988 ). Evidence McCarthy systems meaning . thebrain . NRture - 430 , 334 , 428 -made : Parry in the namingof naturalandman McKenna , P., & , R. (1994 ). Category specificity . : Normative data from adults and children Rth Rbilit Rtion 4 255 281 . , , objects Neuropsychological -specific : Warrington : A singlecase McKenna , P., & , E. K. (1978 ). Category naming preservation . Rnd 41 571 574 . , , , , Journal study of Neurology Neurosurgery Psychi Rtry . (1980 : Warrington . Journal McKenna , P., & , E. K , ). Testingfor nominal of Neurology dysphasia - 788 . , Rnd , 43 , 781 Rtry Neurosurgery Psychi McNeil . (1994 : Warrington of propernames . Neuro , J. E., Opolotti , L, & , E. K ). Theaccessibility . , 32 , 193 208 psychologia ., & : Caramazza of lexicaland sublexical Miceli , G., Capasso ,R , A. (1994 ). The interaction proin reading and . 32 317 333 . cesses , , , writing repetitionNeuropsychologia : Caramazza of dissociation in i, M. C., Nocentini , V., & Miceli , G., Silver , A (1988 ). Patterns and of . nouns and verbs 2 351 358 . , , Aphasioiogy comprehension production i , M. C., Villa : Caramazza for the agrammatics Miceli , G., Silver , G. P., & , A. (1984 ). On thebasis - 220 mainverbs . Coria, 20 . , 207 difficultyin producing . F. (1994 with spared Miozzo : Cappa actionnaming dueto a , A., Soardi , S., & ). Pureanomia , S - 1109 . Neuropsychologia . left temporal lesion , 32 , 1101 in the human Nobre : McCarthy inferiortem , G. (1994 , A C., Allison , T., & ). Word recognition . NRturt 372 260 263 . , , porallobe -medial in the anterior Nobre : McCarthy -relatedfield potentials , A C., & , G. (1995 ). Language lobe : D. Effects of word and semantic . Neuroscience 15 1090 , , Journal of temporal type priming . 1098 of language . Journal , G. A (1991 , 11 , 2281 ). Corticalorganization Ojemann of Neuroscience 2287 . . E., Fox : Raichle emission Petersen , S , P. T., Posner , M.IMintun , M., & , M.E. (1988 ). Positron - 589 of the . NRturt . studies cortical of word , 331 , 585 anatomy single tomographic processing : Raichle . (1989 emission Petersen , S. E., Fox , P. T., Posner , M. I., Mintun , M., & , M. E ). Positron studies of the of words . Neuroscitna 1 , , 153 Journal of Cognitive processing single tomographic . 170

294

Tests andClini~ 1Syndromes Neuropsychological

Pietrini , V., Nertempi , P., Vaglia : Ferro , A , Revello - Milona , H.G., Pinna . V., & , F. (1988 ). Recovery from herpes : Selective of sped with simplex encephalitis impainnent Ac semantic categories - 1293 correlation . Journal , Neurosurgery , andPsychiatry . , 51 neuroradiological of Neurology , 1284 Raichle , M. E., Fiez , J. A , Videen , T. 0 ., Macleod , A.-M. K., Pardo : Petersen , J. V., Fox , P. T., & , -related s. E. (1994 in human brainfunctional ). Practice nonmotor learning changes anatomy during . Cerebral Cortu , 4, 8- 26.
: Humphreys Riddoch M . J., & , G. W. (1987 : A caseof ) . Visual object processingin optic aphasia semanticaccess . CognitiveNeuropsychology , 4, 131- 185. agnosia Roelofs , A (1993). Testing a non- decompositional theory of lemma retrieval in speaking : Retrieval of verbs. Cognition , 47, 59- 87. Rosch , E., Mervis, C. B., Gray, W. D., Johnson : Boyes - Braem , D. M ., & , P. (1976). Basicobjects in natural categories . CognitivePsychology , 8, 382- 439. Sacchett : Humphreys , C., & , G. W. (1992). Calling a squirrel a squirrel but a canoea wigwam: A category-specificdeficit for artefadual objects and body parts. CognitiveNeuropsychology , 9, 7386. Sartori, G., & : Job . (1988). The oyster with four legs: A neuropsychological study on the ,R interaction of visual and semanticinformation. Cogni HveNeuropsychologyiS , 105- 132. : Zettin, M . (1988). Generatingproper names Semenza , C., & : A caseof seledive inability . Cogni HveNeuropsychology , 5, 711- 721. : Zettin, M . (1989). Evidence&om aphasiafor the role of proper namesas pure Semenza , C., & . Nature , 342 , 678- 679. referring expressions ' Shallice T. : Kamounis, L D. (1993). Seledive impainnent of retrieving people , ,& s names : A category disorder ? Coriu 29 281 291. , , specific Silveri , M . C., & : Gainotti, G. (1988). Interaction between vision and language incategory HveNeuropsychologyis , 677- 709. spedfic semanticimpainnent. Cogni Small : Gordon, B. (1995). Distributed representationsof semantic , S. L , Hart, J., Nguyen , T., & , 118 , 441- 453. knowledge in the brain. Brain : Vanderwart, M . (1980). A standardizedset of 260 pictures , J. G., & : Nonns for name Snodgrass , image agreement , familiarity, and visual complexity. Journalof &perimental Psychology agreement : HumanLearning and Memory, 6, 174- 215. ' : Szikla Talairach , J., & stiriota.riquedu tilincephale , G. (1967). Atlas d anatomie . Paris : Masson . : Farah , G., & , M . J. (1996). A category-spedfic naming impairment after Tippett , L J., Glosser , 34, 139- 146. temporal lobectomy. Neuropsychologia Tranel, D. (1991). Dissociatedverbal and nonverbal retrieval and learning following left anterior . Brain and Cogni Hon , 15 , 187- 200. temporal damage Tranel, D., Damasio : Damasio . (1995). Double dissociationbetweenovert and covert , H., & , A. R facerecognition . Journalof Cogni HveNeuroscience , 7, 425- 432. Tranel, D., Damasio : Damasio . (1997a , H., & ,A R ). On the neurology of naming. In H. Good: A . Wingfield ( Eds .), Anomia : neuroanatomical and cogni Hve correlates glass & (pp. 65- 90). New York: AcademicPress . Tranel, D., Damasio : Damasio . (1997b , H., & ,A R ). A neural basisfor the retieval of conceptual , 35, 1319- 1327. knowledge. Neuropsychologia Tranel, D., Damasio ., & : Brandt, J. P. (1995). Separateconceptsare retrieved , H., Damasio ,A R &om separateneural systems : Neuroanatomicaland neuropsychological double dissociations . , 21, 1497. for Neuroscience Society

295

Tranel, Damasio : Damasio : The Neural Basisof Lexical Retrieval ,&

-related . J., & : Damasio . (1997 Tra Ml , R ,A R , D., Logan , C. G., Frank ). Explaining category andlexical for CCX \aete entities : of conceptual elects in the retrieval mowledge Operationaliza . of fadors . Neuropsvchologill - 1339 tion andanalysis , 1329 , 35 . (1975 of semantic . Qu , E. K ). The selective Irrlerly ] oumalof memory Warrington impainnent - 657 . , 27 , 635 & pIrimtnt RlPsychology - 196 . .K . (1981 . British , 72 , 175 ,E ). Conaeteword dyslexia ] oumRl of Psychology Warrington . Bmin ., & : McCarthy . A (1983 , 106 , , E. K ,R ). Categoryspedficaccess Warrington dysphasia 859 - 878 . .K ., & : McCarthy . A (1987 of mowledge , E ,R ) . Categories : Furtherfractionations Warrington 11 . . Bmin 110 1273 96. andanattempted , , integration . Brmn . K., & : Shallice , 107 , , E , T. (1984 ). Categoryspedficsemantic Warrington impainnents . 81 .9- 854 with automated . (1993 . P., Mazziotta : CherryS. R Woods ,R , J. C., & ). MRI-PETregistration . Assisted - 546 . ] oumQl , 17 , 536 of Computer Tomography algorithm of pure . S. (1988 class andcontexteffedsin a case : Berndt ,R , LB ., & ). Grammatical Zingeser - 516 . Cognitiw . an : Implications for models of language < Hnia , 5, 473 Neuropsychology production . S. (1990 of nouns andverbsin agrammatism andan < Hnia . : Berndt ,R , LB ., & ). Retrieval Zingeser 39 14 31 .. Brmn andlAng , , ~

296

andClinicalSyndromes a J Tests Neuropsychologic

IV

Applications in Dementia

13

A Modelof Human Based onthe Memory Cellular of theHippocampal Physiology Formation


MichaelE. Hasselmo , BradleyP. Wyble, and ChantalE. Stem

In recent years, the term cognitive psychology has given way to the term cognitive neuroscience , as researchers working on cognitive function embrace a range of new techniques . However , this change in terminology does not yet reflect a substantive change in the focus of cognitive psychology . Most researchers still appear to think in terms of step-by - step flow charts of cognitive function , seeking ever more Anely tuned dissociations in behavior . The current use of brain -damaged subjects and neuroimaging techniques appears primarily to focus on mapping macroscopic brain structures to the superstructure of cognitive psychology . This approach still ignores the vast of neuroscience data, which are primarily obtained at the microscopic majority or molecular level . But effects at the cellular and receptor level clearly playa fundamental role in cognitive function - the range of memory effects of different drugs provides a good example . The work presented here attempts to bridge the gap between macroscopic behavior and microscopic function . In this chapter we present a detailed network model of the hippo campus which allows data on human memory function to be analyzed in terms of physiological and anatomical data at th ~ cellular level . We demonstrate how the model has been used to simulate effects of drugs such as scopolamine and diazepam on memory function . In addition , we show how the model can be used to understand the effect of lesions of the hippocampal formation and the progression of cortical neuropathology in Alzheimer ' s disease. Computational modeling is essential to bridging this gap . Here , detailed computational models of cortical networks are utilized to model human memory function . Though it addresses the same topic , this work contrasts with extensive previous work on models of human memory function . Previous memory modelers have frequently stated the eventual need to confront biological realism ( Murdock, 1982 ), but most published work has utilized simple and abstract mathematical operations to model memory : Shiffrin , 1981 ; Metcalfe , 1993 ; Murdock & : Kahana, 1993 ). (Raaijmakers & These are sophisticated models which are useful for interpretation of behavioral data and specific hypotheses framed on a more intuitive level , but they are not true functional models . They do not address the actual brain

of this, the constraints applied to involved in memory. Because mechanisms these models have been constraints of simplicity rather than biological validity . But it appearsthat the complex and often paradoxical data on human memory performancecan only be explained by multiple overlapping systems . It will be important to constrain such multiple system models with biological data. Here we attempt to approachhuman memory function from a new perspective by drawing on extensive researchand modeling of the hippocampalformation. The wealth of data concerning the hippocampal formation provides an excellent opportunity to examine how theories developed at the behavioral level could be implementedusing models incorporating detailed anatomical and physiological data. Behavioral data from humans suggest that the hippocampusmediatesthe learning of memories for specific episodes(Scoville : Milner , 1958; Corkin, 1984; Zola-Morgan, Cohen, & : Milner , 1957; Penfield& & : Amaral, 1986). This includeswork on patients with lesionsdue to surgical removal of large portions of the medial temporal lobe, such as patient HiM . : Milner , 1957; Corkin, 1984), as well as patients with damage (Scoville & restricted to subregionsof the hippocampus , as in patient RiB. (Zola-Morgan et al., 1986). Patients with damage to the hippocampus appear capable of retaining information for a short period without distraction commonly displaying normal behavior in tests such as Digit Span (Corkin, 1984). In addition, they appear to retain most episodic information obtained prior : Squire to their lesion (Zola-Morgan, Cohen, & , 1983; Zola-Morgan et al., - the inability to 1986). However, they display severeanterogradeamnesia . For example form new long-term memories , in the paired-associates portion HiM . shows no recall of unrelated of the Wechsler Memory Scale , patient : Milner , 1957). When tested three times sequentially (Scoville & paired associates on a list of ten unrelated paired associates , patient RiB. only recalled 3.7 pairs out of 30 (Zola-Morgan et al., 1986) and other hippocampal : Press showed zero recall (Squire amnesiacs , 1990), while controls , Amaral, & recalled correct associationsover twenty times. In tests of the free recall of ten words from the middle of a fifteen-word list, patient RiB. only recalled 10 percent of the words, whereas controls recalled about 40 percent (Graf, : Mandler, 1984). Similar striking differencesbetween controls and ,& Squire patients with hippocampallesions appear in tests of the free recall of information from a story- commonly a subtest of the Wechsler Memory Scale : Milner , 1958 ; Zola-Morgan et al., 1986). These data support (Penfield & the specific significance of hippocampal subregions in storage of verbal information. Animal data further support a role in episodic memory. Similar to human amnesiacs , monkeys with hippocampal lesions can retain information for short periods (under 10 seconds ), but perform poorly at longer delays in : Squire task a delayed nonrnatchto sample , (Alvarez-Royo, Zola-Morgan , & 1992). lesions affecting hippocampal function have been shown to impair " " : Harrison, 1989). Behavioral formation of snapshot memories (Gaffan &

300

in Dementia Applications

data from rats performing different taskshave led to a number of theories of hippocampalfunction. Sutherlandand Rudy have proposed that the hippocampusis required for learning configural representationsof stimuli which cannot be representedby simple links between preexisting concepts(Suther land & : Rudy, 1989; Rudy & : Sutherland , 1989). Eichenbaumand colleagues have proposed that the hippocampus is essentialfor encoding of relations between multiple stimuli, but emphasizethat theserelational representations canbe usedin a more flexible mannerfor subsequent behavior than configural : Buckingham , 1990; Eichenbaum : representations (Eichenbaum& , Otto , & Cohen , 1994). Extensivework has focusedon the role of the hippocampusin performing tasks requiring formation of spatial representations(McNaughton & : Morris , 1987). These theories have been discussedin the context of neural network models , but they have not been explicitly simulated in a . detailed manner . Here, theories about the function of individual hippocampal cell populations are evaluated in the context of a full network simulation of hippocampalformation with specificbehavioral function. This network simulation hasbeen tested for the capability of forming rapid representations of sequentially " " presented episodic information. The model can be used to directly simulate cortical process es involved in performance of standard human tasks , including the cued recall of paired associates memory , and the free recall of lists of words. However, in contrast to previous models of human memory, the physiological detail of this model allows it to addressquestions at a biological level. Thesequestionsinclude patterns of physiological activity which might underlie memory function, and the cellular basis for drug effects on human memory. Here we illustrate the possible cellular mechanisms for the storage impairments associatedwith the drugs scopolamine and diazepam . In addition, this model can be used to addresspossible mechanisms for the initiation and progressionof cortical neuropathologic changes in Alzheimer' s disease .

SIMPLE OVERVIEWOFNEURAL NETWORKS


Most neural network papers assumetoo much knowledge on the part of their reader . Here a brief overview of biologically realistic neural network models is presented , basedon the neural networks coursetaught by Michael Hasselmoat Harvard. Those who know this material can skip to the next . section To understandthe function of the model presentedhere , is it important to understandtwo basic types of neural network models : associativememories and self-organizing systems . These two basic functions appearin many subregions of the hippocampussimulation, and form the basis for the function of the simulation. These neural network models always involve interacting populations of neurons , representedby the circles in figure 13.1. The populations of

301

Hasselmo I: Stem: A Model of Human Memory , Wyble, &

--- ..J ~ -Input 9 . y ~ Input


B Region 1

1 Region 00eeO0e-

Learning

Recall

Input

2 Region

p p Input 1

Figure 13.1 A. Associative memory function allows activity in region 1 to evoke an associated pattern in region 2. During learning (left), external input setspatterns of activity in region 1 and region 2 (shading representsactive neurons ). Hebbian synaptic modification strengthens ). During recall connectionsbetween active units (thick lines representstrengthenedconnections (right ), external input sets the pattern of activity in region 1. Activity spreadsacrossstrengthened connections to activate the previously associatedpattern of activity in region 2. B, Self. organization allows patterns of activity in region 1 to evoke compressedor nonoverlapping representationsin region 2. Dynamics are the sameduring learning and recall. External input is presentedto region 1 but not region 2. The spreadof activity acrossinitially random synaptic . Connectionsare then strengthened connectionsfrom region 1 to region 2 activatesspecificneuroos between active neuronsin region 1 and active neuronsin region 2, and weakenedbetween of individual patterns . This setsup region 2 representations active neuronsand inactive neurons of patterns in region 1. or classes

neurons interact via distributed connections or synapses, as represented by the lines connecting the circles in figure 13.1. Activity spreads &om one neuron to another via these synapses using an activation rule . Learning involves changing the strength of connections . In the model presented here, learning will always use a Hebbian rule , that is, connections between two active neurons will be strengthened .

302

Applications in Dementia

. .

Input 2

Associative

Memory

Fundion

In associativememory models , learning allows activity in one set of neurons to evoke an associatedpattern of activity in another set of neurons . If the two sets of neurons are in the sameregion, this is referred to as " autoassociative " memory. If the two sets of neurons are in different regions, this is " referred to as "heteroassociative memory. These models have two phaseswith different dynamics : a learning phase and a recall phase . During learning , external input setsthe pattern of activity in each set of neurons , as shown on the left in figure 13.1A. Hebbian synaptic modification strengthens connections between active units, as shown by the thicker lines. Note that external input is the predominant influenceon postsynapticactivity during learning. , external input is given to one set of neurons During recall , but not to the other, as shown on the right in figure 13.1A. The activity then spreads acrosspreviously strengthenedconnectionsto activate the previously associated . pattern of neurons This simple description applies to a wide range of models , including linear associativememories(Anderson, 1972), which involve a single step of recall , and attractor neural networks ( Hopfield , 1984), which involve settling over many stepsof activity during recall. Note that external input is the predominant influence on postsynaptic activity during learning, but NOT during recall. Thus, associativememory function requiresexternal mechanisms for suppressingtransmissionat modifiable synapsesduring learning (Hasselmo , 1994 , 1995). In the model presented here , this can be provided by cholinergic suppressionof synaptic transmissionduring learning. This could also be provided by suppressionof synaptic transmissionby activation of y-aminobutyric acid type B (GABA . ) : Hasselmo : Hasselmo , 1994; Wallenstein & , 1997) or other receptors(Tang & . receptors
Self - Organization In self-organizing systems, learning allows patterns of activity in one set of neurons to freely develop compressed or altered representations in another region , without any external guidance . These networks are often referred to " " as competitive neural networks , and are used to model the development of feature detectors and topographic maps. During learning, external input sets different patterns of activity in one set of neurons , but not in the second set of neurons , as shown in figure 13.18. This means that the second set of neurons responds on the basis of the spread of activity across the connections , which usually starts with a random distribution of strengths . Hebbian synaptic modification strengthens connections between active neurons in region 1 and neurons which happened to be active in region 2.

303

Hasselmo : Stern : A Modelof Human , Wyble ,& Memory

For self-organization , there must be some basis for competition between different units for synaptic weight . This can be obtained by weakening connections betweenan active unit and an inactive unit. This canalso be obtained normalization of synaptic strength, so the total sum of connectionstrength by remainsthe same . This meansthat strengthening of one connection results in . In most self-organizing systems , learning weakening of other connections . However, in the occurs during multiple presentations of many patterns model presentedhere , sequentialself-organization in responseto single patterns is obtained. , external input setsparticular patterns of activity in region 1. During recall . The region The activity spreadsinto region 2 via strengthenedconnections 2 activity can be interpreted as a categorization or compressedrepresentation of the patterns in region 1. For example , , in the model presentedhere highly overlapping patterns in region 1 representing a stream of episodic with nonoverlapping patterns with smallernumbers memorieswill be associated involved of active neuronsin region 2. Note that in contrast to synapses in associativememory function, synapsesinvolved in self-organization are the predominant influence on postsynaptic activity during learning. This might require suppressionof the postsynaptic influence of other intrinsic . synapses The self-organization function describedhere applies to a wide range of different models , most of which addressthe formation of feature detectors and topographic mapsin the primary visual cortex (Miller , Kelier, 8t Stryker, 1989). A HIPPOCAMPAL MODEL OF FREERECALL AND CUED RECALL The network simulation of hippocampal function consists of interacting populations of neurons representing the activity of a number of different . Self-organization of input connections allows the hippocampal subregions formation of nonoverlapping , compressedrepresentationsof individual episodic memories . Associative memory function within and between regions allows input cues or context cues to evoke the stored patterns and send . them back to neocortical structures
Local Networks of Hippocampal

PY !' amidal Cells and Intemeurons

This network simulation used simplified representationsof individual neurons designedto mimic basic properties of hippocampalpyramidal cells and intemeurons (Hasselmo& Stem, 1997; Hasselmo , 1995). , Schnell , & Barkal Eachunit in the model had a variable representingmembranepotential (relative to a resting potential of - 60 to - 70 mV ). As in real neurons , this variable Different models used showed passivedecay back to resting potential. . In spiking neurons either spiking neuronsor continuous output neurons , the

304

in Dementia Applications

membranepotential would vary continuously until it reacheda firing threshold , at which point a spike would be generated , and the potential would be reset to a lower value. Continuous output neurons were similar except that when the membrane potential reached firing threshold, the unit output would increaselinearly in proportion to the amount by which activation exceededthe threshold.

Contrast with ConnectionistModels


in any model, but the simplifications present in Simplificationsare necessary this model differ consider ably from most of the connectionistmodelsyou will find in this book. The models presentedhere can be more directly mappedto the neurophysiology of cortical regions. In contrast , many models still have unrealistic features . Three unrealistic features common to physiologically connectionist models are (1) use of an explicit error signal to train the many network; (2) use of sigmoid input-output functions; and (3) use of both positive and negative connectionsarising from a single unit. 1. Unrealistic supervised learning rules. Most connectionist models use - that is, supervised learning techniques techniques in which learning is regulated by an explicit comparisonbetween actual output and desired output . The predominant technique is backpropagation of error. The explicit . Biologically computation of error has not been demonstratedin cortical structures realistic models commonly use learning rules dependent upon information available locally at individual synapses . For example , physiodata from the formation the existence of Heblogical hippocampal support bian synaptic modification (Kelso : Brown, 1986; Wigstrom, , Ganong, & Gustafsson , Huang, et al., 1986). 2. Unrealistic sigmoid input-output functions. The sigmoid input-output function is used in most connectionist models of neocortex (Cohen & : Ser van-Schreiber , 1992) and hippocampus (Burgess : O 'Keefe 1994 , Recce , & , ; Gluck & : Myers, 1993 : O 'Reilly, 1995; Myers ; McClelland, Mc Naughton, & & : Gluck, 1994; Schmajuk& : Di Carlo, 1992). In models with sigmoid inputoutput functions, the activity of individual units often gets very large, but the total output is constrainedat a particular value (usually 1.0). However, cortical neurons almost never fire at the maximum firing rate dictated by intrinsic properties . Neurons can fire briefly at rates up to 200 Hz if driven very hard by current injection, but in awake behaving animals , the firing rates never go above 40 Hz in hippocampus(Ranck : , 1973; Wiener, Paul ,& Eichenbaum : Eichenbaum , 1989; Otto & , 1992) or 100 Hz in neocortex (Hasselmo , Rolls, Baylis, et al., 1989). The shapeof the sigmoid function is also wrong . The threshold in a sigmoid function is defined as the half-maximum , and the output of the neuron rises gradually below threshold as well. In contrast , intracellular recording from cortical pyramidal cells in brain slice : Hasselmo , 1994) shows that for the range of firing preparations (Barkal &

305

Hasselmo : Stem: A Model of Human Memory , Wyble, &

& : Eichenbaum , 1995), most neurons frequenciesobserved in vivo (Schoenbaum have a threshold linear input- output function. 3. Unrealistic combination of excitatory and inhibitory connections arising from individual units. In most connectionist models , individual units give : rise to both positive and negative connections with other units (Gluck & & & : 1994 et al. McClelland 1995 1993 Gluck , ; , ; ; Schmajuk : Myers Myers, functions with small it easier to obtain Di Carlo, 1992). This makes complex numbers of units. However, it contrasts with physiological data showing a strict division between excitatory and inhibitory neurons in the cortex. The fast excitatory transmitter, glutamate , is usually releasedfrom the axon terminals never release the fast inhibitory which cells cortical of , pyramidal transmitter, GABA . In contrast , GABA is usually releasedfrom the axon terminals . Thus, there of cortical interneurons , which never releaseglutamate fast of cells a strict division to be excitatory potentials, and causing appears those causingfast inhibitory potentials. Co-releaseof transmitterswith other has been observed , but these modulatory substances modulatory substances . appearto have slower effects
Connedivib

BetweenDifferent

: Subregions

The basiccircuit usedto representthe activity in eachhippocampalsubregion is summarizedin figure 13.2A. The network simulation of the hippocampal formation contained five subregionswith different functions, including entorhinal cortex layers II and III , the dentate gyrus, hippocampal region CAJ , region CA 1, and entorhinal cortex layer IV . Local circuits representingthese subregionswere linked together to representfunctional interactions within the hippocampal network. The anatomical structure of the hippocampus is : Witter , 1989, for anatomical summarized in figure 13.28 (see Amaral & review), along with brief descriptionsof the proposed function of individual . hippocampalsubregions
Dynamics

of Individual Subregions

Episodic memory function within the simulation required specific functional . Here the function of individual capabilitieswithin the hippocampalsubregions in detail. The discussed is strength of individual connections subregions in Hasselmo & Stem is describedin more detail (1997). Rapid Sequential Self- Organization in Dentate Gyrus Self organization of excitatory perforant path connections in the dentate gyrus formed new of each sequentiallypresentedpattern. The manner in which representations this occurredinvolved the initial activation of a subsetof dentate gyrus neurons due to the random initial strength of connections from the entorhinal cortex. Neurons which were active then developed stronger connections

306

in Dementia Applications

A
? ' ? ' . IJ Ai w

of a localcircuitof the hippocampal model . Excitatory Figure 13.2 (A). Connectivity neurons with membrane external and receive potentialai receive Ai input excitatoryinput from units within the regionandin otherregions via the connectivity matrixWij. These unitsalsoreceive via the connectivity matrixHiJ . Inhibitoryinterneurons input from inhibitoryinterneurons with membrane via connectivity matrixW~ and potentialhk receive input from excitatoryneurons receive via the connectivity matrixH'kI . (8) Proposed input from inhibitoryneurons functionof individualanatomical connections within the hippocampal formation . 1. Fibers of the perforant entorhinal cortexlayers II andIII with the dentate . These pathconnect gyrus undergo rapidselforganization to form new representations of patterns to entorhinal cortex presented sequentially . 2. The dentate to regionCAJ via the mossyfibers . These connections gyrusprojects pass the sparse on to regionCAJ for autoassociative . 3. Longitudinal representations association storage fibersconnect cellswithin regionCAJ, mediating autoassociative and pyramidal storage recallof CAJ patterns of activity . 4. The Schaffer collaterals connect regionCAJ with region CA1 andmediate heteroassociative andrecallof associations between storage activity in CAJ and the self - organized formedby entorhinal representations input to CA1 . s. Perforant path connections alsoenterregionCA1 from the entorhinal cortex . These self - organization undergo to form newrepresentations of entorhinal cortexinput for comparison with recallfrom CAJ. 6. backfrom regionCA1 enterlayerIV of the entorhinal Projections cortex , eitherdirectlyor via the subiculum . Thesestoreassociations between CA1 activity and entorhinal cortexactivity , in CA1 to activate the associated in entorhinal allowingrepresentations cortex activity patterns CA1 canin8uence layerIV. 7. Region , eitherdirectlyor via connections activity in the medialseptum with the lateralseptum , allowinga mismatch between recalland input to increase between recallandinput to decrease (ACh ), anda match acetylcholine ACh. 8. Themedial septum (andthe verticallimb of the diagonal bandof Broca ) provides modulation to all cholinergic . hippocampal subregions

IJ a

with active neurons in entorhinal cortex , while connections with other inactive entorhinal cortex neurons became weaker due to the synaptic decay regulated by postsynaptic activity . In addition , connections between active entorhinal cortex neurons and inactive dentate gyrus neurons became weaker due to synaptic decay regulated by presynaptic activity . This caused a more selective pattern of connectivity between the active entorhinal cortex neurons and the neurons in the dentate gyrus forming a representation of this pattern . At the same time as this occurred , synapses from the active inhibitory interneuron in the dentate gyrus to the active neurons in the dentate gyrus were also strengthened . This ensured that subsequently it would be more difficult to activate the dentate gyrus units which were members of this representation , ensuring that subsequent patterns of activity in entorhinal cortex would evoke different patterns of activity in the dentate gyrus . Without this feature , the network tended to lump all entorhinal cortex patterns into the same representation within the dentate gyrus .

Autoassodative Function of Recurrent Collaterals in CA3 Once individual dentate gyrus units became sufficiently active, the activity passed -to -one connections representing the mossy fibers and activated one along individual CAJ units. Initially , when cholinergic suppressionwas present , the pattern of activity was primarily determinedby this afferent input rather than the strong recurrent excitation. This allowed selectivestrengthening of between the neuronsactivated by the dentate gyrus. As these synapses synapses becamestronger, the activity in region CA3 increased , resulting in greater output to region CAt . As noted below, increasedactivity in region CAt causeda decreasein cholinergic modulation throughout the network. This decreased modulation resulted in greater recurrent excitation in region CAJ , such that the network entered a stable fixed point attractor pattern representingthat stored memory. This stable fixed point attractor was the main driving force for recall within the network. For example , when a cue pattern missing certain components of the previously stored input pattern was presented , it would more weakly activate individual neurons in the dentate gyrus. However, if the , input was sufficient to activate a subcomponentof the CAJ representation then the recurrent excitation would greatly strengthen this activity , pushing CA3 activity into the previously stored fixed point attractor state. Thus, consider ably weaker input could elicit the same amount of CA3 output, thereby driving activity in region CAt and entorhinal cortex layer IV. Rapid Sequential Self- Organization in Region CA 1 At the same time as sequentialself-organization took place in the dentate gyrus, it also took . In this region, similar learning place in region CAt of the hippocampus dynamicsat excitatory and inhibitory connectionsmediatedthe formation of

308

in Dementia Applications

new representations . However, the consider ably weaker input from the entorhinal cortex meant that this input alone could not strongly activate region CAl , but required conjoint input from region CAJ . During learning , the perforant path and Schaffercollateral inputs would interact to form selforganized , the pattern of activity representationsin region CAl . During recall in CAJ would elicit an associatedpattern of activity in CAl , which would usually match the pattern of direct input from entorhinal cortex. When this match was sufficient , it would suppresscholinergic modulation and allow recall to be strongly driven by region CA3 output . Heteroassodative Function of Schaffer Collaterals to CAt As mentioned above , the output from region CA3 activated region CAl . During the initial learning of a novel pattern, the random initial connectivity of the Schaffercollaterals causeda distributed pattern of activity in region CAl , which would interact with the perforant path input to form a new selforganized . Subsequently , during recall, the perforant path representation would have a influence on activity in CAl . However, initially input stronger if the Schaffercollateral input matched the perforant path input sufficiently to causereduction of cholinergic modulation, then the cholinergic suppression of synaptic transmissionat the Schaffercollaterals would be removed. This allowed Schaffer collateral activity to dominate within region CAl , driving neurons strongly which had previously been associatedwith the particular activity pattern in region CA3. Heteroassodative Feedback from Region CAIto the Entorhinal Cortex Initially , during learning of a new pattern, the feedback from by cholinergic modulation, allowing adivity region CAt would be suppressed in entorhinal cortex layer IV to be dominated by the input coming via the one-to-one connections from layer II. This allowed storage of associations between the new pattern of activity in region CAt and the simultaneous , during recall, pattern of activity in entorhinal cortex. Subsequently cholinergic modulation would be reduced (as describedin the previous section ), thereby removing the suppression of synaptic transmission at the feedbackconnectionsfrom region CAt . This allowed activity in region CAt to effectively reactivatethe previously associated pattern of activity in entorhinal cortex layer IV. Feedback Regulation of Cholinergic Modulation from the Medial Septum The level of activity in region CAt determined the level of chol. This allowed the network to inergic modulation from the medial septum respond initially to patterns with dynamics set by strong cholinergic modulation , suppressing the autoassociative and heteroassociative function of . This prevented associativerecall at the Schaffercollaterals excitatory connections from interfering with effective self-organization of perforant path

309

Hasselmo : Stem: A Model of Human Memory , Wyble, &

input to region CAI . In addition, this prevented previously stored representations from interfering with the storageof new associations at excitatory recurrent connectionsin region CA3, at the Schaffercollaterals connecting region CA3 and region CA 1, and at the feedbackconnections from region CAI to entorhinal cortex layer IV. When a sufficient level of activity in , the cholinergic modulation would be reduced region CA 1 was obtained , transmission to mediate associative recall at allowing strong synaptic synapses in region CAJ , region CAI , and entorhinal cortex layer IV. Thus, the feedbackregulation of cholinergic modulation sets appropriate dynamics for , though the cholinergic modulation goes through the learning and recall sametransition from high to low during presentationof eachinput pattern.
Modification

of Synapses

Modifiable synapsesin the model all utilized a similar Hebbian learning rule dependent upon postsynaptic and presynaptic activity , in keeping with experimentalevidenceon the Hebbiannature of long-term potentiation in the , 1979; Mc Naughton, Douglas, & Goddard, hippocampus (Levy & Steward 1978; Kelso et al., 1986; Wigstrom et al., 1986). This samelearning rule provided self-organization of perforant path synapses from entorhinal cortex to dentate gyrus and region CA 1, and associativememory function at the longitudinal associationfibers in region CA3 and the Schaffercollaterals projecting to region CAI . The different functional properties did not arise from differencesin the learning rule, but from the fact that cholinergic suppression of synaptic transmission during learning resulted in associative memory function, while the absenceof suppression during learning allowed selforganization (Hasselmo , 1994 , 1995). The simulation of both learning and recall within this model contrastswith other theoretical work on the hippoan initial pattern of connectivity and then focuseson campuswhich assumes the dynamicsof recall (Treves & Rolls, 1994). The rapid sequential self-organization of perforant path synapsesin the dentate gyrus and region CAI of the model also required decay of synaptic strength regulatedby the amount of pre- and postsynaptic activity . Decay of synaptic strength can be taken as a representationof the phenomenon of , as describedin experimental preparations(Levy, Collong-term depression bert, & Desmond , 1990). In most simulations , weights were clipped at specific values to maintain them within the region of stable attractor dynamics . Thus, the strength of modifiable connections did not exceed parameters termed Wmax . The connections from inhibitory interneurons in the dentate gyrus and region CAI were also modified. This selectively increasedinhibition to units which respondedstrongly to an individual input pattern, making it more difficult for these sameunits to be activated by other patterns and ensuring their selectivity primarily for patterns closely matching the . pattern to which they first responded

310

Applications in Dementia

Feedback Regulation

of Cholinergic

Modulation

The total output from units in region CA1 determined the level of cholinergic modulation within both region CA3 and region CA1. Increased cholinergic modulation from the medial region CA1 output causeddecreased . This is consistent with experimental of neurons via activation inhibitory septum evidencesuggestingthat activity in region CA1 and region CA3 can inhibit activity in the medial septum , and thereby downregulate cholinergic modulation ( McLennan& Miller , 1974). When levels of cholinergic modulation were high, there was strong suppressionof synaptic transmissionat the excitatory recurrent synapsesin CA3, as demonstrated experimentally (Hasselmo et al., 1995), at the Schaffer collaterals projecting from region CA3 to CA1 ( Valentino & Dingledine, 1981; Hasselmo & Schnell , 1994), and at projections from CA1 to entorhinal cortex layer IV. This prevented the spread of activity due to previous learning from interfering with selforganization . When levels of cholinergic modulation were decreased , the recall associative transmission was increased of , allowing strength synaptic . Simulationsalso included the suppressionof inhibitory synaptic to dominate transmission(Pitier & Alger , 1992) in dentate gyrus, region CAJ , and region CA1, and the suppressionof excitatory input to inhibitory intemeurons in region CA1. The cholinergic enhancementof excitatory synaptic modifica tion (Burgard & Sarvey , 1995) was represented , 1990; Hasselmo & Barkal . Finally, cholinergic modulation causeddirect at all modifiable connections depolarization of inhibitory and excitatory neurons in all regions (Barkal & Hasselmo , 1982; Madison & Nicoll , 1984). , 1994; Benardo& Prince

'V FUNCTION MEMOR IN THE MODEL


Basic features of human memory function were replicated in the model , using patterns of input in entorhinal cortex layer II as representations of the input during behavior , and using patterns of activity in entorhinal cortex layer IV as the output from the network . The function was analyzed in network models of the hippo campus using either spiking neurons or continuous output neurons . The spiking neuron simulations presented here have been demonstrated to perform cued recall for individual paired associates. The continuous output neuron simulations have been used to analyze learning and free recall of lists of items . Modeling Cued Recall in a Spiking Model of Hippocampal Function

Learning and recall of paired associates was demonstrated in the network model of the hippo campus constructed from simulations of spiking neurons . An example of the function of this network is shown in 6gure 13.3. This 6gure shows the activity of neurons in different regions of the simulations

311

: Stem: A Model of Human Memory Hasselmo , Wyble, &

Figure 13.3 Network model of memory function in the hippocampususing spiking neurons . Activity of eachof a population of neuronsin eachsubregionof the model is representedby the width of black lines. Spiking appearsas wide lines in each trace . Leaming occurs when cholinergic modulation in the network is set at high levels (A Ch modulation is plotted on the righthand side). Input is presentedas a pattern of activity .in entorhinal cortex layer II. This forms in the dentate gyrus, which are passedalong to subsequent rapid self-organized representations regions. T ransrnissiondelays between the different regions bridge the gap between sequential , allowing activity on one cycle to be associatedwith activity on a subsequent gamma oscillations cycle during learning. Recalloccurs when cholinergic modulation in the network is at a low level. Pres ~ntation of a degraded version of the input pattern activates the previously formed representationin dentate gyrus. The spreadof activity through the full loop reactivates the full representationin entorhinal cortex layer II.

(plotted horizontally) for eachtime step of the simulation (plotted vertically ). Eachblack squarerepresentsa spike generatedby the simulation. As seen in the figure, a particular paired associatewas presented to the model in the form of a distribution of spiking activity in entorhinal cortex . layer II. Four neurons are activated on two consecutivegamma oscillations This input was presentedin the presenceof cholinergic modulation (activity of cholinergic input is plotted on the far right ). The activity then spreads through the network, activating a subsetof dentate gyrus neurons , and then consecutivelyactivating neurons in region CA3, region CAI , and entorhinal cortex layer IV. Synaptic modification during this activity results in storage of this self-organized representationin dentate gyrus, and associativememory function in CA3, CA 1, and entorhinal cortex layer IV. Later, the recall of the associationis tested by presenting a degradedversion of the input pattern to entorhinal cortex layer II, during the absenceof cholinergic modulation. In the absenceof cholinergic modulation, new representations are not formed, but the stronger synaptic transmissiondue to the absenceof suppressionallows strong activation of representationsin CA3, CAI , and entorhinal cortex layer IV . The activity completes the full

312

in Dementia Applications

IV

layer

cortex

CA

gyrus
II

>

>

JIg

rl

'

. .

10

'

&

'

ri

"

Entor

"

'

'

'

"

"

Region

C'

RegIon

'

. ,

po

rr

ntate

: . t ; , : . . . : : : ; , : * , . ! , : t : : : . . . , i : ~ . . : . " . ~ ' : . . . . . : . , ~ . . : . : . ' . . Entorhinal layer cortex , , , i I : ! . t , ! 1 ;

, :

~ . 3 ~
120 40 80

Gamma

cycle and reactivatesthe full learnedrepresentationin entorhinal cortex layer II. This constitutes recall of the full pattern (and also allows further rehearsal of the pattern for better storage ). In this framework, recall from the hippocampus requires transmission through the full trisynaptic loop, with the direct connectionsfrom entorhinal cortex to regions CAJ and CA 1 providing a test of the validity of recall at eachcycle. In this framework, recurrent connections within region CA3 are not the driving force for attractor , but could instead provide a mechanismfor bridging between different dynamics , or predictions loops, constituting links between different memories about the next state of the environment. Linking Episodic Memory Function to Physiological Data on Eledroencephalographic and Single -Unit Recordings Use of a spiking model of the hippocampusin evaluating episodic memory function allows this function to be discussedwith direct reference to the availablephysiological data on hippocampalformation. Ext l\sive recordings have been obtained from the hippocampus in awake behaving rats. Most recordings have been performed during exploration of the environment : Mc Naughton, 1993), but responseshave also been , 1973; Wilson & (Ranck , 1973), analyzed during performance of various behavioral tasks (Ranck nonmatch to : discrimination and delayed sample(Otto & including olfactory Eichenbaum , 1992). In addition, considerableresearchhas focused on oscillatory (EEG ). Lowdynamics within the hippocampalelectroencephalogram frequency oscillations in the theta range (3 to 10 Hz) have been described extensively within the hippocampusduring exploration of the environment : Colom, 1993). And higher-frequency gamma oscillations , 1973; Bland & (Ranck : Buszaki have also been described (Chrobak & , 1991). The firing of interneurons has been known to correlate with theta oscillations , leading " " Luo : Fox, 1973 Stewart to their initial designation as theta cells (Ranck , ,& , ; to 1992). More recently, spiking of pyramidal cells has been proposed maintain ' : Recce , 1993). specificphaserelations with the theta rhythm (O Keefe & This allows the pattern of neuronal spike generation in the model to be discussed with regard to the pattern observed during electrophysiological experimentsrecording single-unit spiking activity or EEGfield potentials. Recently it has been proposed that hippocampal neurons should fire within individual gammacycles , but only once during each theta cycle (Lisman & : Idiart, 1995). This proposal may partly have arisen from the general . But these report of very low average firing rates in hippocampal neurons neurons actually show brief periods of high frequency activity . In fact, pyramidal " cells in the hippocampus are commonly referred to as complex" cells becauseof their tendency to emit bursts of action potentials spike : Miles, 1991). due to an interaction of intrinsic conductances(Traub & These bursts could be ideal for driving the activity in CA3 into a particular attractor state corresponding to an individual memory since bursts result in

313

t Slem: A Model of Human Memory Hasselmo . Wyble. &

repetitive output from a single neuron at intervals of only a few milliseconds , allowing several spikes within a single gamma oscillation. Inaddition to the generationof bursts , the neuronswill fire at up to 30 to 40 Hz for brief periods of time, suggestingrepetitive action within a single theta oscillation . This type of activity would be necessaryif reverberatory activity occursacrossgammaoscillations within the hippocampus . The model presentedabove uses different time periods with cholinergic modulation present or absent . These changesin modulatory dynamics are with particularly important regard to the theta rhythm . It has been suggested that cholinergic modulation is important for induction of theta rhythm oscillations . This might be due to sustained cholinergic modulation inducing theta dynamics . If this is the case , it is unclear how both learning and recall can occur within the network, since the model presentedabove requiresdifferent dynamics for learning vs. recall. With regard to cholinergic modulation and theta rhythms, there are different possibilities . 1. Cholinergic activity follows the theta cycle. The levels of acetylcholine (A Ch) within the hippocampusmay rise and fall at the samerate as the theta rhythm, giving alternating cycles of learning and recall dynamics . Cholinergic neurons in the medial septum appearto fire in synchrony with theta cycles (Stewart et al., 1992), but the levels of A Ch in the hippocampushave not been me" sured at such a fast time base . Microdialysis measurements of A Ch levels are usually performed over a period of ten minutes (Inglis, Day, & : Fibiger, 1994), preventing sucha measurement . If this is the case , feedback of modulation the match of recall and regulation dependentupon input may set the amplitude of oscillations rather than changing a static level. However, recent data from this laboratory suggests that cholinergic effects are too slow for this- taking over a second to reach maximum amplitude, and decaying over many seconds(Fehlauand Hasselmo , unpublisheddata). 2. Cholinergic activity is tonic and sustainedduring the full period of theta . If this is the case , then changes between learning and recall dynamics must dynamics , since animals are clearly depend on other neuromodulators performing both learning and recall while they explore their environment. It is possible that cholinergic modulation sets a general activation tone during exploration, and then removal of this modulation results in sharpwave activity and consolidation , as proposed by Chrobak and Buszaki(1991). The continuum between learning and recall dynamics could instead be due to the activation of GABAB receptors on presynaptic terminals ( Wallensteinand Hasselmo , 1997). Activation of these receptors causessuppressionof synaptic transmissionspecificto connectionsarising from region CA3 pyramidal cells (Ault & : Nadler, 1982; Colbert & : Levy, 1992). The levels of GABA within the hippocampus clearly must change in time with the theta oscillation -linked to theta , since inhibitory interneurons are commonly phase oscillations (Skaggs et al., 1996; Fox et al., 1986). If this is the case , each

314

in Dementia Applications

theta oscillation could constitute a shift between dynamics dominated by feedbackand dynamicsdominated by afferent input . Modeling Paired Associate Learning in a Continuous Firing Rate

Model

More extensive analysis of the storage of multiple patterns has been performed using a continuous firing rate representation of individual hippo. This model hasbeen used to replicate the learning and cued campalneurons recall of a list of paired associates , as well as the learning and &ee recall of a list of single items. For testing paired-associatememory, activity in the full hippocampalnetwork was evaluated during sequential presentation of a series of highly overlapping activity patterns in the entorhinal cortex for 400 time steps each , followed by incomplete cues for each pattern. (The network activity .) This could be consideredanalogousto was reset to zero between patterns , followed by presentation of one presentation of a list of paired associates word &om eachpair for cued recall. Incomplete pattern cueswere chosenso . as to be unambiguous The network demonstratedthe capability to store nine highly overlapping input patterns presentedsequentially to the region representing entorhinal cortex layers II and III during steps 0 to 3600, followed by recall of these highly overlapping input patterns in response to incomplete cues during steps 3600 to 7200. The input and output of the simulation is shown in figure 13.4. Note that the output on the right side of the figure (entorhinal cortex layer IV ) shows almost identical responseto the nine incomplete cues as to the nine complete input patterns. This sequentialstorage and recall of highly overlapping patterns is interpreted as the capability to store episodic . memories Figure 13.4 also shows the activity of individual neuronsin the full hippo campal network during learning of the nine highly overlapping patterns of input activity . Each new pattern evoked a unique pattern of activity in the dentate gyrus, which was passedon to region CAJ , where it was reinforced by recurrent excitation. The interaction of perforant path input &om entorhinal cortex and Schaffercollateral input &om region CA 1 then activated a with the pattern unique representationin region CAl , which was associated IV. in cortex of activity entorhinal layer Sequentiallypresentedpatterns activated different representations , despiteoverlap between the stored patterns. With no externally induced change in network dynamics , the network demonstratesthe capability to recall stored patterns in responseto incomplete cues , as shown in steps 3600 to 7200 in figure 13.4. Though the cues contain only 50 percent of the usual input activity , they sufficiently activate the previously stored representationsin the dentate gyrus such that region CAJ enters the previously stored fixed point attractor state. This then activates the region CAI representation and the associatedfull pattern in

315

: A Modelof Human : Stern Hasselmo , Wyble ,& Memory

self sclf = o rg / " " uto ~ " = / ~ I : ~ U = = 0 bete ~ asso bo ~ - "


SnJA8
ident;it)' mab"lx

xa1JO

Figure 13.4 Activity in the full hippocampalnetwork during paired-associatememory function (learning of sequentially presentedoverlapping input patterns ). Horizontal width of black lines representsthe output value of eachneuron in the network during the full time period of the simulation . This includes ten excitatory neurons and one inhibitory neuron in entorhinal cortex layer II, thirty -one excitatory neurons and one inhibitory neuron in the dentate gyrus, region CAJ , and region CAl , and ten excitatory neurons and one inhibitory neuron in entorhinal cortex , activity is shown for one inhibitory and one cholinergic basalforebrain layer IV. In addition neuron regulating cholinergic modulation. During the initial 3600 time steps , nine different patterns of activity are induced in entorhinal cortex layers II and III for 400 stepseach . In response to eachpattern, the network rapidly fonns a self- organizedrepresentationof the pattern in the dentate gyrus and region CAl . The dentate gyrus pattern is autoassociated in region CAJ , and associations are stored between the activity in CAJ and CAl and between CAl and entorhinal cortex layer IV. During steps 3600 to 7200, incomplete versions of each input pattern are presented to entorhinal cortex layers II and III. These evoke the previously formed self- organized in dentate gyrus, which induce recall activity in region CAJ. This recall activity representations drives region CAl along with the perforant path input, and the CAl activity inducesrecall of previously stored patterns in entorhinal cortex layer IV. This allows the network to respond with activity in layer IV on steps3600 to 7200 which matchesthe patterns learnedon steps0 to 3600.

316

in Dementia Applications

'

II

II

II

II

II

II

II

"

iden \ity mab' lx

~ Self-org

~ hetcro assoc

AI

, '

'

&

tl

II

'

'

"

'

'

"

"

II

II

'

'

'

'

' '

II

"

"

II

II

'

"

'

'

'

'

"

'

'

'

II

'

'

"

"

'

"

'

'

'

3O

'

Number

II

'

'

Neuron

'

"

'

II

II

'

'

' ' '

II

II

'

II

II

"

"

"

"

. ' . I I I , , , It I , 1 I . . I " I I , I 1 : , ~ r I I : ' 11 II I I II " rrlrlll ' III 1 J rillr I I 8 ~ .

'

entorhinal cortex layer IV . Comparison of the patterns in entorhinal cortex layer IV during steps 3600 to 7200 with those during steps 0 to 3600 show that the network has effectively recalled the nine stored patterns . However , note that the representations of each new pattern become more and more sparse as learning continues , resulting in weaker storage of memories stored later in the sequence. This detailed model of the hippo campus suggests a possible mechanism for the primacy effect observed in list learning . Separate simulations tested the capability of the network to respond to the incomplete cues when presented intermittently with the novel patterns to be stored . The network effectively set appropriate dynamics for storage and recall of the individual patterns without external regulation of network dynamics . Extensive simulations demonstrated the breakdown of memory function in . Interference during recall could occur when the high degree of certain cases overlap between stored patterns resulted in recall cues being ambiguous (a cue with only two active lines would often result in recall of all patterns in which those two lines were active ). This cue ambiguity could be overcome by ensuring that the cue only matched one stored pattern , or by adding an extra unambiguous input line . Interference during learning could also occur when a novel pattern matched a previously stored pattern by 75 percent of the input lines. In this case , the two patterns would be chunked together with units which initially formed the representation of the first pattern , and elements of either pattern would recall both patterns . This almost never occurred with patterns overlapping by 50 percent or less. Finally , as additional overlapping patterns were stored , the number of neurons available for forming new representations gradually decreased, resulting in sparser and , later patterns did not develop a stable sparser representations . In many cases representation , and could not be recalled with degraded cues.

Modeling FreeRecallin a ContinuousFiring Rate Model


. In contrast to pairedModeling free recall presents additional problems , associatelearning, which has a distinct cue for each stored representation with all items should be free recall requires that a generic context associated able to elicit recall of the individual items. This presentsa difficult problem, but free recall was obtained in the model due to the attractor dynamicspresent in region CA3. The structure of the model presentedabove was modified to allow separate attractor dynamics for context and item information. This required separate subpopulationsin each hippocampal subregion to respond to the item information or to the context information. The dentate gyrus, CAJ , and CA1 were all subdivided , with individual feedback interneurons for each subdivision mediating separateattractor dynamicsfor item and context. The function of this network is shown in figure 13.5. In region CA3, weak synaptic connectionsbetween the item and context information allowed weak

317

: Stem: A Model of Human Memory Hasselmo , Wyble , &

Input

Context

1 2 3 4

I I I I I I

" ' 11111 . 11111 I " I ,!1 IIII I III I I I I II I I I III I II [ 'N 'N 'N

I I II hi I In (. 1 I 'N 11111 . 1 " " 'r " 1 I '" I II Itr 1111 II . Entorhinal Dentate Region Region Entorhinal cortex CA3 CAI cortex gyrus layerII layerIV is shown and Figure13.5 Theactivityof different hippocampal subregions during storage free recall of a setof fourinputpatterns . Each has been divided intosubregions for item region information and for context interneuron for each information , withanindividual subregion inhibitory . There areweaker feedback connections between thecontext anditemsubregions in CA . 3. Thenetwork fonns self -organized of each of theinputpatterns region rapid representations . During recall of theinputis presented , onlythecontext to ento during learning portion rhinal cortex II. Thisactivates theassociated of context in the dentate layer representation gyrus andin region CA . 3. Thecontext enters anattrador state in region CA . 3 which representation then drives recall of individual diHerent in theother items of region CA . 3. Each attrador portion state in thenetwork is tenninated the intrinsic of individual neurons . by adaptation properties Thepersistence of adaptation currents withintheneurons thesame from prevents pattern being adivated on sequential . Notethatthenetwork recalls three of thestored cycles sequentially items before a repetition of thefullsequence . starting associationsto be formed between the context and each of the individual stored items. Subsequently , presentation of context only as a cue would evoke the attractor state associatedwith context. This attractor state would then activate one of the item attractor states . Transitions between different attractor states in the free-recall model uti lized an explicit model of the intrinsic adaptation properties of hippocampal of the buildup of intracellular pyramidal cells. The model contained representations calcium during firing activity , and of the calcium - dependentpotassium current mediating adaptation in each pyramidal cell. This adaptation would build up during a period of activity , and eventually prevent further activity , terminating a particular attractor state. Persistenceof the intracellular calcium levels in a previously recalled item attractor state would result in hyperpolarization, preventing previously recalleditems from recurring on each recall cycle, but often items would appear repetitively over a longer cycle. As shown in figure 13.5, this network was capableof learning a set of items and then recalling individual items from the list.

" ' ( ! UI I I II I 'I" r I I II I I I I ( , r I " 'I ' II" "" 'IIHII " Itl " t " " 'I" 'II - 11 '" I It ' r " ' I r II "" ' " 1" l II IIUIIIIII " ' " ' II 111 t r Ilv

' I I I MI

I II ' III1 , I I III .

( 1111 . 1 I 1111 hi - I

~ 3 ~
.

318

in Dementia Applications

Modeling

the Effects of Scopolamine and Diazepam on Free Recall

This biologically detailed model of hippocampal memory function allows analysis of how the effect of drugs at a cellular level influences memory function. Drugs such as scopolamineand diazepamhave been demonstrated to impair the learning of new information. Thesedrugs do not affect the free recall of lists of words learned before injection of the drug, but they impair the recall of lists of words learned while under the influence of the drug : Mewaldt , 1975; Peterson , 1977). The effects appear to be specific (Ghoneim & . For example to long-term storage rather than short-term storage , the is numbers forward and backward a list of to ( ) ability digit span not repeat recall of . In addition immediate 1978 affected by scopolamine(Drachman , , ) the last few words in a list of words (the recencyeffect Murdock, 1962) is son, : Grove-White , 1973; Frith, Richard not affectedby scopolamine(Crow & Samuel , et al., 1984). Thus, the recall of material stored without distraction is unaffected . However, once there is a period of distraction, even short-term . Persistent activity in preexisting representationsin the storage is affected neocortex and hippocampus may mediate the retention of information for short periods without distraction. In contrast , retention of information for a distraction seconds of may require newly formed representations during period in the hippocampus . Retention of information for brief periods during : Peterson distraction is testedby the Brown-Petersontask (Peterson& , 1959), in which consonanttrigrams are presentedbefore a distractor task. Scopol amine impairs recall of information in the Brown-Peterson task (Caine , . al. Butters & : 1986 , Janowsky , ) Weingartner, Ludlow, et , 1981; Beatty, The effect of scopolamineand diazepamon memory function is due to effects at specific receptors within the brain. Scopolamine(also known as . That is, it selectively blocks hyocine) is a muscariniccholinergic antagonist of A at a the effects of Ch subtype cholinergic receptor which is sensitive to muscarine(in contrast to nicotinic receptors ). The effects of A Ch at muscari nic receptors within cortical structures have been described extensively in : Phillis, 1963; Benardo brain slice and whole-animal preparations( Krnjevic & Hasselmo & : Bower, 1992; HasMadison & : Nicoll 1984 & : Prince 1982 ; , ; , : Hasselmo : Schnell selmo & , these , 1994). As describedabove , 1994; Barkal & effectshave been implementedin the model, allowing analysisof how their blockadeimpairs memory function. ) is an agonist at the benzodiazepine receptor site of the Diazepam( Valium GABAA receptor complex. The GABAA receptor protein contains a channel of chloride ions. When activated by GABA , this channel which allows passage . Activation of the benzodia has an inhibitory effect on cortical neurons , zepine receptor site by diazepamenhancesthe conductivity of the channel of effects The enhancement effect. inhibitory giving a stronger inhibitory within the model can be used to model the effects of diazepamon memory function.

319

Hasselrno : Stem: A Model of Human Memory , Wyble, &

Context

cortex layer II

gyrus

CA3

CA I

cortex layer IV

. The network shown Figure 13.6 Storageof patterns for free recall impaired by scopolamine in figure 13.5 was tested with decreased levels of cholinergic modulation during learning. The learning of new representationswas impaired due to the absenceof cholinergic depolarization and the slowed rate of synaptic modification . Only weak activity is elicited in responseto the individual patterns presentedto entorhinal cortex layer II. During recall , the context information does not elicit any item information within the network .

13 .6 shows that partial blockade of cholinergic effects withinthe Figure modelinterferes with the learning of newinformation for freerecall , but doesnot interfere with the freerecallof information learned beforethe administration of scopolamine . In thisexample of learning is , theimpairment dueto the lossof cholinergic of neurons which , depolarization prevents them fromreaching sufficient for synaptic modification activitylevels during . In addition when neurons do show some , even , theabsence learning activity of thecholinergic enhancement of synaptic modification slows theprocess of . These effects new but because learning prevent learning , cholinergic suppression of synaptic transmission is alsoabsent canrespond , the network to previously learned . effectively patterns Theeffects of diazepam differconsider of scopolamine ---ablyfromthose the strength of inhibition the network . theyinvolveincreasing throughout - preventing However of diazepam is similar , theneteffect sufficient activity for learning of new input patterns the recallof previously , but not preventing stored . patterns 'S A BREAKDOWN IN NETWORK FUNCTION AND ALZHE I ME R DISEASE NEUROPATHOLOGY
If parameters are not balanced properly , the simulation undergoes a breakdown in function referred to as runaway synaptic modification (Hasselmo , 1994 ). Runaway synaptic modification results from synaptic transmission

320

Applications in Dementia

II

( I I r I I I I I I I Entorhinal I . I I I I I I I I I I , I I I Region I I I I ' . I I . . Region I I I II , , I I t , I I

II

"

' I I ' . . . Dentate ' I I I h I I I I I I I r I r ' ' ' I I I II II I II I I II II i ' , I I I I I , I I I I III I I I I I I I ' I , I Entorhinal I I I I II I I I I II II

= 'C 0 i~ 3
1 2 3 4

~ 3 ~

during synaptic modification causingan exponential growth in the strength of individual connections and the number of strengthened connections . In associativememory function, runaway synaptic modification can occur if the cholinergic suppressionof synaptic transmission is insufficient to prevent recall of previously stored associationsfrom interfering with the storage of new associations , 1994). In self-organizing systems (Hasselmo , runaway synaptic modification can occur if normalization or synaptic decay is not sufficiently . strong enough to prevent activity from spreadingto all neurons modification has been Runaway synaptic proposed to underlie the initiation and progression of neuropathologic changesin Alzheimer' s disease , 1994). This has been used to explain the apparent selective sensitivity (Hasselmo of the hippocampal formation to the initial appearanceof neuro, and the eventual high density of neurofibrillary tangles fibrillary tangles found in this region (Braak & : Braak of its role in sequential , 1991). Because , and its rapid rate of synaptic storage of a large number of similar patterns modification, the hippocampusis more sensitive to the phenomenonof runaway synaptic modification. The greatly increaseddemandson the metabolism and axonal transport mechanismsof a neuron strengthening a large number of individual synapsescould result in the excessiveproduction of ' the molecular markers of Alzheimer s disease , including the tau protein, which makesup a large componentof neurofibrillary tangles . In the hippocampalsimulation , runaway synaptic modification can appear initially at the connections from entorhinal cortex layer II to the dentate gyrus if there is a reduction in the decay rate of connectionsto the dentate , the breakdown in function can spread to other gyrus. Once this occurs a to , owing regions sharp increasein the amount of overlap between different stored representations . Thus, the breakdown can spreadto the feedback connectionsfrom region CA 1 to entorhinal cortex layer IV , and via a similar mechanism can spreadon into other neocortical regions. ' If this breakdown in function reflectsthe actual progressionof Alzheimer s disease , then the model can be used to generatepredictions about the neuropsychologica ' characteristics of Alzheimer s disease . In particular, it suggests ' that at an early stage , Alzheimer s disease should be characterizedby a . This prediction is being currently greater sensitivity to proactive interference tested using a modification of the Brown-Peterson task focused on . As the diseaseprogress es , it will lead to a complete proactive interference breakdown in episodic memory function, which is an extensively described ' . Other neuropsychological features of Alz feature of Alzheimer s disease ' heimers diseasehave been discussedin the initial description of this theory , 1994). (Hasselmo ROLE OF ACETYLCHOLINE IN NEOCORTICAL FUNCTION

Hasselmo : Stem: A Model of Human Memory , Wyble, &

as well, since neocortical structureshave also been demonstratedto have a , 1991), and the capacity for Hebbian synaptic modification ( Rauschecker same effects of cholinergic modulation have been described within neocortical structures(Hasselmo& Cekic , 1995; Kmjevic and Phillis, 1963). The common anatomical and physiological features of neocortical structures suggest that all neocortical regions share basic operating principles , 1988). Studies by neuropsychologistsof the breakdown of cortical (Shepherd function can both contribute to and benefit from researchfocused on . understanding the common processingcharacteristicsof cortical structures Researchwith physiologically realistic models is essential to this goal, because the wide range of cortical models which utilize connectionist architectures suffer from the failings described above - they do not sufficiently conform to physiological evidence to be mapped to the brain. As an alternative , some of the circuit properties describedhere might provide a means for obtaining the characteristicsof error backpropagation within a more physiologically realistic framework. Here the properties of a model of neocortical function are described , , self-organizing interactions , and drawing on the associativeattractor states matching function describedabove. The basicunit of cortical function will be a conditional self-sustaining attractor state. Excitatory interactions within a local cortical region (suchas a column) allow this self-sustainedactivity , similar to the sustainedactivity in the hippocampalmodel shown in figures 13.4, 13.5, and 13.6. These local excitatory loops will interact with other local circuits via self-organized connections . The self-sustainedattractor within a local circuit is conditional upon input from other regions , just as hippocampalactivity is conditional upon a match between CA3 input and entorhinal cortex input to region CAI . This follows simple rules: 1. Input from another region initiates the loop, but must match the loop " " recall in order for the loop to becomeself-sustaining . 2. If there is no subsequent input, the loop will persist in activity for a period of time (working memory), but will eventually decay due to adaptation . 3. The loop will send output to other cortical regions via self-organized connections . Thesewill initiate loops in other areas , which can interact either or with the . initiator negatively positively 4. It is possible for a loop to activate another loop which gives negative feedback , shutting off the first loop . This is not due to long-distanceinhibitory connections , but to excitatory connections which do not match , and act activation of local . inhibition thereby primarily through 5. Positive feedback can result in metaloops which can progress through in which activation of step 2 loops shuts off step 1 loops. metasequences 6. When a particular input does not activate any previously established loop, then high levels of cholinergic modulation allow formation of a new

322

Applications in Dementia

attractor state within the most active of columns . Formation of new attrac tors can occur across multiple columns , forming new global attractors . 7. In cases when multiple active loops are incompatible , the network will switch between different states until it finds a state that matches the incoming sensory constraints , or until new attractors are formed which can reconcile the conflicting attractor states. 8. This network can interact with a network containing representations of episodic memories , in order to arrive at a final consistent state without requiring repeated presentations of the individual conditions . Further development of this model of neocortical function will allow cognitive function to be directly linked to the dynamics of cortical physiology .

SUMMARY AND DISCUSSION


The simulation of the hippocampus presented here combines functional hypothesesabout individual hippocampalsubregionsinto a network capable of learning and recalling episodic memories . Simulations shown here have addressed the type of function required for paired-associate learning and free recall in tests of human memory function, and have demonstratedhow the blockade of cholinergic effects by scopolaminemay impair learning of new information. Within the hippocampal model, rapid self- organization of perforant path synapsesin the dentate gyrus and region CA1 results in a pattern of synaptic connectivity which elicits sparse restricted patterns of activity . These patterns can be construed as representationsof specific episodes (specific patterns of afferent input from the entorhinal cortex). These representationsare distinct for each memory, allowing effective retrieval of missing elementsof highly overlapping patterns. The model presented here is not incompatible with another memory model focusing on different dynamicsduring theta rhythms and sharp waves , 1989). In that theory, removal of synaptic suppressionwith much (Buzsaki lower levels of A Ch allows large-scalespreadof excitatory activity from the , as supported by physiological hippocampusback into neocortical structures data (Chrobak & : Buzsaki , 1994). That model focuseson contrasting storage of information in the hippocampus and consolidation of that information, that is, the transfer from hippocampusto neocortical structures(seeMcClelland et al., 1995). The model presentedhere focuseson the relative influence of previously stored representations or external input on the representations activated within the hippocampus by novel vs. familiar information. Consolidation has addedin this network (Hasselmoet al. 1996). The simulation presentedhere provides a detailed , explicit description of how hippocampalformation could mediate the formation of episodic memories . The loss of a structure which rapidly forms sparse representations of sequential input could result in the profound anterograde and partial

323

Hasselmo : Stem: A Model of Human Memory , Wyble, &

: retrograde amnesiaobserved in subjects such as patient HiM . (Scoville & Milner , 1957) and patient RiB. (lola -Morgan et al., 1986). The sparing of remote episodic memory can be addressedin a simulation in which the stored hippocampalrepresentations gradually causethe formation of discrete neocortical representationsvia feedbackconnections(Hasselmoet al., 1996; McClelland et al., 1995 : Rolls, 1994). Further expansion of this ; Treves & model to include neocortical structures will allow simulation of a broader range of human memory data, including serial position effects in immediate free recall. This simulation draws together hypothesesabout individual hippocampal , allowing analysisof the dynamic interaction of different regions. subregions The dentate gyrus has frequently been proposed to form sparse , distributed of in the entorhinal cortex (McNaughrepresentations overlapping patterns ton, 1991; Treves & : Rolls, 1994; O 'Reilly & : McClelland, 1994). However, is enhancedby sequentiallearning of separateself-organized representations the use of a novel mechanismpresentedhere , in which inhibitory connections were strengthenedin order to prevent previously formed representations from dominating learning in responseto new representations . Region CA3 hasfrequently beenproposed to have autoassociative memory function : Morris , 1987; Levy et al., 1990; Treves & : Rolls, 1994), but (Mc Naughton & this requires regulation of the strong influence of excitatory feedback . Here feedbackinhibition is used to limit network activity (Hasselmoet al., 1995). In addition, excitatory feedback is prevented from interfering with new learning due to cholinergic suppressionof synaptic transmissionat recurrent , similar to the mechanismproposed previously for pirlform cortex synapses : Bower, 1992; Hasselmo , Anderson, & (Hasselmo , 1993, 1994). The Schaffer collateralshave beenproposed to undergo either self-organization (Treves & : Rolls, 1994) or heteroassociativememory function (Mc Naught on , 1991). The relative amount of these two functions dependsupon how strongly the Schaffercollateralsinfluenceactivity in region CA1. If the Schaffercollaterals dominate postsynapticactivity in region CA1 during learning, then they will predominantly undergo self-organization, as proposed earlier. This must be assumedif no mechanismfor modulation of synaptic transmissionis incorporated in the models . However, with the cholinergic suppressionof synaptic transmission at the Schaffer collaterals , the perforant path input can more strongly influence CA1 activity during learning, allowing heteroassociativememory function. This is necessary if the function of region CA1 is to provide a comparisonbetween recall activity produced by region CA3 and direct input from entorhinal cortex, as proposed by some researchers : Buckingham (Eichenbaum& , 1990). This comparison function will work most effectively if combined input from both entorhinal cortex and CA3 is necessaryto causeregion CA 1 neurons to fire. In the simulations presented here , a comparisonfunction of this type plays an important role in allowing cholinergic modulation to set appropriate dynamicsfor learning and recall.

324

Applications in Dementia

ACKNOWLEDGMENTS
This chapteris supported by an Office of Naval Research Young Investigator -93- 1-0S9S AwardNOOO14 andan NIMH awardR29MHS2732 -O1 .

REFERENCES
Alvarez-Royo, P., lola -Morgan, 5., & : Squire , L. R. (1992). Impainnent of long-tenn memory and of short tenn in : A responseto sparing memory monkeys with medial temporal lobe lesions Brain Research , 52, 1- 5. Ringo. Behavioural Amaral, D. G., & : Witter , M . P. (1989). The 3- dimensionalorganization of the hippocampal formation- A review of anatomicaldata. Neuroscience , 31, 571- 591. Anderson , J. A . (1972). A simple neural network generating an interactive memory. MathematiC /ll Bioscienct S , 14 , 197- 220. Ault , B., & : Nadler, J. V. (1982). Baclofenselectively inhibits transmissionat synapsesmade by axons of CA3 pyramidal cells in the hippocampalslice . Journalof Phannacology and Erptrimental , 223 , 291- 297. Therapeutics Barkal : Hasselmo , E., & , M . E. (1994). Modulation of the input/ output function of rat pirifonn cortex pyramidal cells . Journalof Neurophysiology , 72 , 644- 658. : Janowsky , N., & , D. S. (1986). Patterns of memory failure after scopolBeatty, W. W., Butters amine treatment : Implications for cholinergic hypotheses of dementia . Behavioraland Neural , 45, 196- 211. Biology Benardo : PrinceD . A (1982). Ionic mechanisms , L. 5., & of cholinergic exdtation in mammalian , 249 , 333- 344. hippocampalpyramidal cells. Brain Research : Colom, L. V . (1993). Extrinsic and intrinsic properties underlying oscillation and Bland , B. H., & in Neurobiology , 41, 157- 208. synchrony in limbic cortex. Progress : Braak Braak , J., & , E. (1991). Neuropathologicalstaging of Alzheimer-related changes . Acta Neu82 239259. , , ropathologia : Sarvey , E. C., & , J. M . (1990). Muscarinic receptor activation fadlitates the induction of Burgard Letters , 116 long-tenn potentiation (LTP) in the rat dentate gyrus. Neuroscience , 34- 39. ' : O Keefe , N., Recce , M ., & , J. (1994). A model of hippocampalfunction. Neural Networks Burgess , 7, 1065- 1081. Buzsaki , G. (1989). A two- stage model of memory trace formation: A role for " noisy" brain states . Neuroscience , 31, 551- 570. Caine , E. D., Weingartner , H., Ludlow, C. L., Cudahy, E. A ., & : Wehry, S. (1981). Qualitative -inducedamnesia . Psychopharmacology analysisof scopolamine (Berlin , 74- 80. ), 74 : Buzsaki Chrobak , J. J., & , G. (1994 ). Selectiveactivation of deep layer retrohippocampalneurons . Journalof Neuroscience , 14 , 6160- 6170. during hippocampalsharpwaves -Schreiber Cohen : Servan , J. D., & , D. (1992). Context, cortex and dopamine - A connectionist . PsychologiC /ll Review , 99, 45- 77. approachto behavior and biology in schizophrenia Colbert, C. M ., & : Levy, W. B. (1992 ). Eledrophysiological and phannacologicalcharacterization of perforant path synapses in CAl : Mediation by glutamatereceptors . Journalof Neurophysiology , 68, 1- 7. Corkin, S. (1984). Lasting consequences of bilateral medial temporal lobectomy: Clinical course and experimentalfindings in HiM . Seminars in Neurology , 4, 249- 259.

325

Hasselmo : Stem : A Modelof Human , Wyble ,& Memory

-White Crow : Grove of the learning deficitfollowinghyoscine , T. j ., & , I. G. (1973 ). An analysis - 327 administration to man . British Rrmi U : oiogy . , 49 , 322 Journal of Ph Dradunan andmemory . In M. A. Upton , D. A (1978 ). Central , A. Di Mascio , system cholinergic . F. Killam (Eds & : K .), Psychopharmacology : A generation - 662). New York : (pp. 651 of progress Raven Press . : Buckingham . (1990 : Experiment Eichenbaum . H., & ,J ). Studieson hippocampal , processing . In M. Gabriel & : J. Moore(Eds .), Learning andcomputational : Foundations neuroscience theoryandmodel -2 nthoorks .31 . (pp. 171 , MA: MIT Press ). Cambridge of adaptive : Cohen of thehippocampal Eichenbaum . H., Otto, T., & , N. J. (1994 ). Two functional components . Behavioral andBrain Sciences . - 518 , 17 , 449 memory system Fox ., & : Ranck thetarhythmandthe 6ring of neurons , S. E., Wolfson ,S , J. B. (1986 ). Hippocampal - 508 in walkingandurethane anaesthetized rats .E BrainResearch . , 62 , 495 .rptrimental son ., Samuel ., & : McKenna Frith , C. D., Richard , J. T. E of , M., Crow , T. J , P. J. (1984 ). Theeffects intravenous andhyoscine . Quarterly memory Journal of Erptrimental diazepam Psychology uponhuman - 144 . , A, Human , 36 , 133 Psychology Erptrimental Gaffan : HarrisonS . (1989 andscene : EHects of fornix transection , D., & ). Place memory memory - 212 in themonkey . Erptrimental Brain Research . , 74 , 202 : Mewaldt of diazepam and scopolamine on storage Ghoneim , M. M., & , S. P. (1975 ). EHects , - 262 retrieval andorganization esin memory . Psychophannacologia . , 44 , 257 process : Myers mediation of stimulus : AcomGluck , M. A., & , C. E. (1993 ). Hippocampal representation - 516 . Hippoamtpus . , 3, 491 theory putational Graf ., & : Mandler that amnesic do not , P. A., Squire , L. R , G. (1984 ). The information patients - 178 . Journal . Learning . , Memory , andCognition , 10 , 164 of E Psychology .rptrimental forget Hasselmo and learningin a corticalassociative . Neural , M. E. (1993 ). Acetylcholine memory 5 32 44. , , Computation Hasselmo . (1994 modification in models : Implications of cortex for , M. E ). Runaway synaptic - 40. Alzheimer ' s disease . Neural Nthoorks ,7 , 13 Hasselmo and corticalfunction : Modelingthe physiological , M. E. (1995 ). Neuromodulation basis of behavior . Behavioural Brain Research , 65 , 1- 27. Hasselmo : Bower modulation of cortical , M. E., Anderson , B. P., & , J. M. (1992 ). Cholinergic associative . function .30 - 1246 . , 67 , 12 Journal memory of Neurophysiology Hasselmo : Barkal modulation of activity - dependent , M. E., & , E. (1995 ). Cholinergic synaptic in rat pirlformcortex . Journal . - 6604 , 15 (10 ), 6592 of Neuroscience plasticity Hasselmo : Bower . M. (1992 to intrinsicnot afferent , M. E., & ,J ). Cholinergic suppression speafic - 1229 in rat pirlform(olfactory fibersynapses . Journal . ) cortex , 67 , 1222 of Neurophysiology Hasselmo : Cekic of synaptic transmission , M. E., & , M. (1996 ). Suppression mayallowcombination of associative feedback and self - organizing feed forward connections in the neocortex . - 161 BrainResearch Bthavioural . , 79 , 153 Hasselmo : Nalwa -centered , M. E., RoDs , E. T., Baylis , G. C., & , V. (1989 ). Object by encoding -selective face neurons in the cortexin the superior sulcus of the monkey . Erptrimental temporal - 429 Brain Research . , 75 , 417 Hasselmo : SchneD of the cholinergic of synaptic , M. E., & , E. (1994 ). Laminar selectivity suppression in rat hippocampal transmission region CA1: Computational modelingand brain slice - 3914 . Journal . , 14 , 3898 of Neurosamce physiology

326

in Dementia Applications

Hasselmo : Barkal , M. E., Schnell , E., & andrecallat excitatory , E. (1995 ). Learning recurrent synapses and cholinergic modulation in hippocampal , 15 region CAJ. Journal of Neuroscience (7), 5249 - 5262 . Hasselmo ., & : Stem , M. E , C. E. (1997 : A simulation ). linking LTPto networkfunction of episodic in the hippocampal -termpoten formation . In M. Baudry & : J. L Davis(Eds .), Long memory - 324 Ha Hon , Vol. 3 (pp. 293 : MIT Press . ). Cambridge , MA Hasselmo ., & : Wyble , M. E , B. P. (1998 in a networkmodelof the ). Freerecallandrecognition : Simulating effects of scopolamine on human . Behavioral Brain hippo campus memoryfunction Research , 89 , 1- 34. Hesselmo : Wallenstein , M. E., Wyble , B. P., & andretrievalof episodic , G. V. (1996 ). Encoding memories : Roleof cholinergic andGAB in the hippo . Hippo , 6, Aergicmodulation campus campus - 708 693 . with graded , J. J. (1984 havecollective ). Neurons Hopfield responses computational properties like those of two -state . Proceedings neurons - 3092 . of theNaHonal , 81 , 3088 Academy of Sciences : Fibiger , F. M., Day , J. C., & , H. C. (1994 release in hippo ). Enhanced Inglis acetylcholine campus andcortexduringthe anticipation andconsumption of a palatable meal . Neuroscience , 62 , 1049 1056 . Kelso . R., Ganong A H., & : Brown ,S in the hippo , T. H. (1986 . Proceedings ). Hebbian synapses campus - 5330 . , 83 , 5326 of theNaHonal Academy of Sciences -sensitive ., & : Phillis ,K , J. W. (1963 cellsin the cerebral Kmjevic ). Acetylcholine cortex . Journal of 166 296327 . , , Physiology : Desmond , W. B., Colbert , C. M., & , N. L (1990 esof neurons Levy ). Elemental adaptive process andsynapses : A statistical . In M. A. Gluck& : D. E. Rumelhart .), / computational (Eds perspective - 236 Neuroscience andconnec Honist : Erblaum , (pp. 187 . , NJ ). Hillsdale theory : Steward , W. B., & asassociative , O. (1979 elements in thehippocampal ). Synapses Levy memory - 245 formation . Brain Research . , 175 , 233 -termmemories ., & : Idiart Usman , J. E of 7 :t 2 short , M. A P. (1995 in oscillatory ). Storage subcycles - 1515 . Science . , 267 , 1512 : Nicoll MadisonD. V., & , R. A. (1984 of rat CA1 pyramidal ). Controlof the repetitive discharge - 331 neuron esin vitro. Journal . , 354 , 319 of Physiology McClelland on, B. L , & : O'Reilly , J. L , McNaught , R. (1995 ). Why are therecomplementary in the hippo andneocortex ? Insights from the success esandfailures of systems learning campus - 457 connectionist models of learning andmemory . Psychological Review . , 102 , 419 : Miller, J. J. (1974 Mclennan , H., & controlof neuronal in the septum ). Thehippocampal discharges of the rat. Journal 237 - 624. , , 607 of Physiology McNaught on , B. L (1991 in hippocampal ). Associative circuits : New evidence patterncompletion - 220 andnewquestions . Brain Research . BrainResearch Reviews . , 16 , 193 McNaught on , B. L , Douglas , R. M., Goddard enhancement in fascia , G. V. (1978 den ). Synaptic - 293 tata : Cooperativity coactive afferents . BrainResearch . , 157 , 277 among McNaught : Morris . G. M. (1987 on , B. L , & ,R enhancement andinformation ). Hippocampal synaptic - 415 within a distributed . Trends in Neurosciences . , 10 , 408 storage memory system Metcalfe . (1993 ,J ). Novelty monitoring , andcontrolin a composite , metacognition holographic associative recallmodel : Implications for Korsakoff amnesia . Psychological Review , 100 , 3- 22. Miller, K. D., Kelier : Stryker , J. B., & , M. P. (1989 columndevelopment ). Oculardominance andsimulation . Science - 615 . , 245 , 605 analysis

327

Hasselmo : Stem : A Modelof Human , Wyble ,& Memory

. ] oumlllof Erperimtntal Murdock . B. B. (1962 , ). Theserialpositioneffectin freerecall Psychology - 488 . 64 , 482 andretrievalof item andassodative information Murdock . B. B. (1982 ). A theoryfor the storage . Psychological Review . - 626 , 89 , 609 : Kahana of thelist-strength effect . ] oumlllof Erperimtntal Murdock . B. B., & , M. J. (1993 ). Analysis Hon - 697 . , Memory , andCogni , 19 , 689 Psychology Learning : Gluck in andhippocampal , C. E., & , M. (1994 , conditioning ). Context Myers rerepresentation - 847 . Behavioral animal Neuroscience . , 108 , 835 learning O'Keefe : Recce between , J., & , M. (1993 ). Phase placeunits and the relationship hippocampal . Hippo EEG thetarhythm 3 317 330 . , , campus O' Reilly : McClelland , R. C., & , J. L. (1994 ). Hippocampal conjunctive encoding , storageand recall : Avoidinga trade -off. Hippo - 682. , 4, 661 campus : Eichenbaum Otto T., & nonmatch , J. (1992 ). Neuronal activity in the hippo duringdelayed campus to sample in rats in recognition : Evidence for hippocampal performance processing - 334 . Hippo . , 2, 323 memory campus : Milner in the hippo Penfield , W., & , B. (1958 ). Memorydeficitproduced by bilaterallesions - 497 zone . Archives andPsychiatry . , 79 , 475 of Neurology campal -termretention Peterson .& : Peterson . ] oumlll of individual verbalitems , L. R , M. J. (1959 ). Short - 198 . , 58 , 193 of& ptrimtntal Psychology Peterson induced failures in man . Psychopharmacology , R. C. (1977 , 52 , ). Scopolamine learning - 289 283 . Pitier : Alger excitation in the rat of GAB , T. A., & , B. E. (1992 ). Cholinergic Aergicinterneurons . oumIlI slice 450 127 142 . , , ] of Physiology hippocampal . M. (1981 . Psychological : Shiffrin of associative , J. G. W., & , R ). Search Raaijrnakers memory Review . , 88 , 93- 134 . B. (1973 of singleneurons in dorsalhippocampal formationand septum in Ranck . J ). Studies unrestrained rats . PartI. Behavioral . Erperimtntal correlates andfiring repertoires , 41 , Neurology - 531 461 . - Hebbsynapses Rauschecker of visualplasticity , J. P. (1991 , NMDA receptors ). Mechanisms - 615 andbeyond . Physiological Reviews . , 71 , 587 : Sutherland formation is necessary for ratsto learn , J. W., & , R. J. (1989 ). Thehippocampal Rudy andremember discriminations . Behavioural BrainResearch . , 34 , 97- 109 configural : DiCarlo , N. A , & , J. J. (1992 , classical ). Stimulus Schmajuk configuration conditioning , and - 305 . Psychological function Review . , 99 , 268 hippocampal : Eichenbaum cortex . 1. Schoenbaum , G., & , H. (1995 ). Information codingin therodentprefrontal in neuron orbitofrontal cortex with that in cortex . OUmlll ] activity of Single compared pyriform - 750 . , 74 , 733 Neurophysiology Scoville . : Milner afterbilateral lesions , W. B., & , B. (1957 ). Lossof recent memory hippocampal - 21. andPsychiatry , 20 , 11 ] oumlllof Neurology Neurosurgery . In MS . Gazzaniga .), , G. M. (1988 (Ed ). A basiccircuit of corticalorganization Shepherd in memory Hves research . Cambridge . , MA: MIT Press Perspec ., McNaught on, B. L., Wilson : Barnes , W. E , M. A., & , C. A (1996 ). Theta phase Skaggs in . neuronal and the of precession hippocampal populations compression themporal sequences - 172 . , 6, 149 Hippo campus

328

Applications in Dementia

. , Amaral : Press , L. R , D. G., & , G. A (1990 resonance ). Magnetic Squire imagingof the hippo formation and nucleidistinguish medialtemporallobe and diencephalic campal mammillary amnesia . Journal - 3111 . , 10 , 3106 of Neurosciena Stewart : Fox , M., Luo , Y. Q., & of atropine , S. E. (1992 on hippocampal ). Effects theta -cellsand -spike - 128 cells . BminRt StRrch . , 591 , 122 complex . J., & : Rudy Sutherland , R , J. W. (1989 association : The role of the hippo ). Configural theory - 144 formation in learning andamnesia . Psychobiology . , 17 , 129 campal , memory : Hasselmo . (1994 , M. E of intrinsicbut not afferent ). Selective fiber Tang , A. C., & suppression - 81. transmission baclofen in the . BminResearch by ) cortex synaptic , 659 , 15 pirlform(olfactory Traub ., & : Miles . (1991 On Rlnehoorks ,R ,R . Cambridge ). Nt Ur . : Cambridge , U.K of thehippoc4lmpus Press . University Treves : Rolls , A, & , E. T. (1994 of the role of the hippo ). Computational in analysis campus - 391 . Hippoc . , 4, 374 memory Rmpus ValentinoR . J., & : Dingledine . (1981 ,R ). Presynaptic in the inhibitory effectof acetylcholine - 792 . }ournRl . ,1 , 784 of Neuroscience hippo campus Wallenstein ., ME . (1997 , C. V., andHasselmo of hippocampal ). GAB Aergicmodulation population : Sequence field and the phase effect . activity learning , place development precession - 408 . , 78 , 393 }ournRl of Neurophysiology WienerS. I., Paul : Eichenbaum , C. A., & andbehavioral , H. (1989 correlates of hippo ). Spatial - 2763 neuronal . }oUrnRl . , 9, 2737 activity campal of Neuroscience -term , H., Gustafsson : Abraham , B., Huang , W. C. (1986 Wigstrom ). Hippocampal , Y.-Y., & long is induced a Herent with intracellu potentiation by pairingsingle volleys Iarly injected depolariz - 319 . Ada Physiologica Scandinavica . , 126 , 317 ing current pulses Wilson : McNaught , M. A , & of the hippocampal on , B. L. (1993 ensemble codefor ). Dynamics - 10S8 . Science . , 261 , 10SS space -Morgan Zola ., & : Squire . (1983 , 5., Cohen , N. J ,L R ). Recallof remoteepisodic memoryin - 500 amnesia . Neuropsychologi . , 21 , 487 R -Morgan Zola ., & : Amaral , 5., Squire , L. R amnesia andthemedial , D. G. (1986 ). Human temporal : Enduring a bilateral lesion limited to field CA1 of the region memoryimpairment following . Journal - 2967 . , 6, 2950 of Neuroscience hippo campus

329

Hasselmo : Stem : A Modelof Human , Wyble ,6 Memory

14

The basal ganglia include the striatum (caudate , putamen , and ventral striaTurn ), substantia nigra (pars reticularis and pars compact a), internal and external globus pallidus, subthalamicnucleus , and thalamicnuclei. Additional limbic and brain stem structures , including the amygdala , red nucleus , and reticular formation, frequently are classified as basal ganglia components . Recent studies have revealed function-specific cortico-subcortical pathways that course in parallel through the basal ganglia system (Alexander, 1995). These neural " circuits" maintain structural and functional segregation in cortex, basal ganglia, and thalamic regions. Additionally , they maintain consistenttopographic mapping throughout striatothalamocorticalpathways. The flow of information is of a feed forward reentrant nature , forming a series of closed loops between cortical and subcortical structures . These pathways have been designatedbroadly as motor (including oculomotor), cognitive , and limbic (Alexander, Crutcher : Delong , 1991; Alexander, Delong, & ,& : Strick , 1986; Schell& : Strick, 1984) (figure 14.1). Eachof these neural circuits is discussed below with regard to structure . , function, and clinical features The basal ganglia playa critical role in many neuropsychological functions , ranging from motor control and planning to attention and emotion. As subcorticalstructures , the basalganglia receive information from widespread cortical sensorimotor and association regions. Information is channeled in parallel through anatomically discrete nuclei and projected back to the frontal cortex through thalamic relays. Within this structure , the basal ganglia form a network of processingcentersvital to the initiation and regulation of goal-directed behavior. Certain biological attributes make the basal ganglia well suited to modeling by neural networks (Beiser : Houk, 1997). , Hua, & These include (a) discrete , serially connectedprocessingcenters ; (b) parallel of information flow c functional of ; ( ) pathways equivalence processingstages for each pathwayd ) convergenceof different types of information at each -regulated ; (e) information flow through reentrant feedback processingstage processingloops; (f ) dynamic adjustment of output basedon changing spa and use of bias or to feedback ; overall ) , . , tiotemporal (g adjust gain output The following discussion emphasizesbackpropagation networks (which are describedin the first chapter of this book), but also presentsseveralother

~ ~ (0 . .> -0 0 -c a ~ ~ CI o I D ~

Limbic
AC MOFC i +-

Cortex

Striatum

Ventral

D rIiJ

< : ~ , ={ ~~ Ventral

I . ca ca E~U I

MD L

Figure 14.1 Three primary basal ganglia pathways. Note dorsal connectivity of motor and : PMA / SMA , primary motor cognitive circuits, and ventral connectivity of limbic circuit. Regions area ; FEF , frontal eye fields; DLPC, dorsal lateral prefrontal cortex ; / supplementarymotor area LOFC, lateral orbitofrontal cortex ; MOFC, medial orbitofrontal cortex; AC, anterior cingulate ; SN, substantianigra ; VT A. ventral tegmentalarea ; VL, ventral lateral; VA. ventral anterior; MD , medial dorsal. Transmittm: DA. dopamine ; glu, glutamate ; GABA . y-aminobutyric acid.

important network models , including self-organizing feature maps and gated ' dipoles . The clinical focus is on motor control and Parkinson s disease (PO ), with additional material on related neuropsychiatric disorders . The chapter concludes with a brief look at future directions in the field of computational modeling and neural network applications to basal ganglia studies.

MOTOR NETWORKS
" " The motor pathways originate in frontal motor regions, including the dorsal and ventral premotor areas , supplementarymotor areas , and cingulate motor areas . Oculomotor circuits originate in cortical frontal and supplementary : Crutcher, 1990). Target regions for motor eye fields (Alexander & are in the , which receivesadditional cortical pathways principally putamen . , as well as brainstem structures input from postcentral somatosensoryareas These pathways are modulated by rich dopaminergic connections originating in the substantianigra (Smith & : Bolam , 1990). Output from the striatum " " " to the pallidum is through two major pathways , the direct and the indirect ' : The direct route involves the neurotransmitters y-aminobutyric acid (GABA ) and substanceP, which act to inhibit output from the striatum to the internal globus pallidus, and from the internal globus pallidus to cortical . Activation of this direct pathway facilitates output areas of the thalamus

332

in Dementia Applications

movement by reducing thalamic inhibition , facilitating tonic excitation of cortical motor regions. The indirect pathway includesan additional pathway &om the striatum to the internal globus pallidus , which projects to the subthalamicnucleus and the cortex. , the thalamus This additional synaptic , connection allows for a tonic inhibitory influence on the thalamus and its cortical projections. Under normal conditions, the direct basal ganglia pathway facilitates thalamic output, while the indirect pathway inhibits it (DeLong, 1990). The basal ganglia projects to the &ontal lobes via thalamocortical pathways , which terminate in regions from which cortical input originated. These connections complete &ontal subcortical circuits of simple and complex motor process es , preserving somatotopic organization throughout the system Hoover & Strick, 1993; Mushiake & Strick, 1993; Parent & Hazrati, ( 1993). Regardless of the original cortical sites , there are specificrostrocaudal target areasin the basal ganglia associatedwith distinct somatotopic areas , but with varying densitiesof dorsal vs. lateral innervation (Strick, Hoover, & Mushiake , 1993). A specific cortico-subcortical input channel may receive , but this information is convergent information &om several cortical sites topographically segregatedat smaller target zones within the basal ganglia themselves(Flaherty & Graybiel, 1993; Selemon& Goldman-Rakic , 1985). PD is the most prevalent movement disorder associated with basalganglia dysfunction, and study of the disorder has provided fundamental information on basal ganglia process es. Discovery of the essentialrole of degenerative changesin dopaminergic substantia nigra cells with resultant loss of input to the striatum, together with demonstrated improvement in PD following dopaminereplacement therapy, hasclarified the neurophysiological basisof the disorder (DeLong, 1990; Graybiel, 1984; Penney& Young, 1983). Primary clinical features of PD include bradykinesia , rigidity , and resting tremor, with associatedfeatures of postural instability , autonomic dysfunction , and oculomotor irregularities (Jankovic& Caine , 1987; Marsden, 1982, 1990). Thesemotor symptoms often are accompanied by impairment in cognitive , 1992; Huber & Cummings , 1992; functioning (Cummings & Benson Mahurin, Feher , Nance , et al., 1993; Mayeux, 1990; Stem & Mayeux, 1986). Detailed studies of motor dysfunction consistently reveal slownessin limb movements , secondary to insufficient scaling of the force of the initial , resulting in severalcorrectiveagonist-antagonistiterations of agonist impulse , Dick, Rothwell, et al., 1986; Evarts , Teravainen velocity (Berardelli , decreasing & Calne , 1981). In addition, there is dampening of repetitive movements such as finger tapping, which progressively decrease in amplitude and velocity over time. PD patients also show disproportionate motor slowing when attempting to executetwo motor programs either simultaneouslyor in , Rothwell, Dick, et al., 1986). These and many sequentialfashion (Benecke other studies of motor function in PD have done much to elucidate the role of the basal ganglia in the initiation and maintenanceof complex motor , 1987). activity (Marsden

333

's Funr Hnn in Parkinson Mahurin : Basal Disease Ganglia

Motor

Control

The basal ganglia are essentialto the planning , and modification , execution of goal-directed motor control. Artificial neural networks have proven to be valuable tools for modeling interactions between various cortical and subcortical structuresvital to the production of such controlled motor behavior. Figure 14.2 illustrates a basic backpropagation network that effectively . Output nodes represent models elementsof a self- correcting motor system motor commandsto specific agonist and antagonist muscle groups, which , and then to the spinal pass from the basal ganglia to cortical motor areas " " . The goal state representsa limb-target intercord and muscle effectors

T T Perceptu8
. A similar architecture can be applied to - correcting motor control system Figure 14.2 Self attention and goal-direded networks as describedin the text. X : , perceptualdata ; Y , inner layer connection ; + ; An , , ; , ( )/ ( ) antagonistD displacement T, time delay. weights As. agonist : Kwan , 1993.) (Adapted from Borrett, Yeap, & 334

in Dementia Applications

= 1 : : : : : : : [

section , and includes both temporal and spatial aspects . Input is acquired visual and , somatosensory , , and reflects through proprioceptive receptors " " " current state information. The " comparator calculatesa temporal integration of output values over units of time, representedby consecutivepasses of data through the network. The comparator function may represent the , or ongoing interaction activity of prefrontal regions, basal ganglia centers ' between cortical and subcortical areas ( Tolkunov, Orlov , 81 : Afanas ev, 1997). Patterns of connection weights dynamically are sampled and selfadjust through backpropagationto minimize the difference value produced the . Movement precision can be enhancedby increasing the by comparator resolution , increasing the quality or quantity of input data, or temporal sharpening the signal-to-noise ratio through lateral competition (Miall , 1995). Descriptions of PD and other movement disorders are basedon the concept of disturbancesin motor programming, in which stored action plans (schemas ) are selected and carried out. However, motor programs per se have no obvious correlates in the activity of the nervous system itself. Borrett, Yeap, and Kwan (1993) designed a network that models the features of a motor program, but without the need for a central controller or . The main feature of this model is an unsupervisedrelaxation programmer of the network into one of three attractor states : Axed , periodic, or chaotic " " ( strange ). The last state is unpredictable , and representstonic disturbance of the system under consideration . The model establish es a dynamic , recurrent network that extends the basic motor control system presentedabove by the addition of nodesfor spatial displacementand a time delay (seefigure 14.2). The time delay allows for reentrant input of the last output state as well as the next-to -last output. Synaptic (connection ) weights are optimized through backpropagation to minimize the mean square error between the ideal limb trajectory (goal state ) and the final output state produced by the network. This network can model both a simple ballistic movement and a repetitive movement about a single joint . Once the network is optimized, a in excitability of the inner layer of nodes can be introduced generalizeddecrease to mimic dopamine depletion in PD. This is accomplishedby shifting the sigmoid transfer function to increasethe threshold value. Final behavior of the network closely parallels that derived from quantitative evaluation of movementsin PD patients (Borrett et al., 1993).
Mechanism ~

Several artificial network models have been developed to demonstrate the critical role of dopaminedepletion in PD (Contreras-Vital & Stelmach , 1995; Kotter & Wickens , 1995). For example , Jamiesonand colleagues(Jamieson , 1991) proposed a network approach to test several dopamine models of movement dysfunction in PD. Eachof severalbase architectureswas modified to simulate different neurophysiological interactions (figure 14.3). The

335

's Mahurin: BasalGanglia Function in Parkinson Disease

. " " , I \ , I , I \ , I
network models developedto test levodopa effectson motor performance Figure 14.3 Pharmacologic ' . A , Dired vs. indirect input to striatum. 8, Two output units. C, in Parkinsons disease , severity; Dys, ; Sev Two cross -connectedoutput units. Mot , motor; Sir, striatum; DA , dopamine . (Adapted in part from Jamieson , 1991.) dyskinesia

networks used a simplified model in which each node representeda neuronal , or treatment outcome. Two input nodes were population, drug effect . Diseasetype : diseasetype and levodopa dosage utilized acrossall models inner node the bias connection an was affecting initially assigned inhibitory a of the the from of aggregatedopamineoutput (hypothetically pars compact an substantia nigra), while levodopa dosage was assigned excitatory connection bias to the samedopaminergicnode. Input and output training data were obtained from a prior study by Mouradian, Juncos , and Fabbrinii (1988), which examined the pharmaco into four were PD. Patients of grouped logical profiles of levodopa treatment : never treated subsets , wearing off, and on/ off (fluctuating). A supervised , stable backpropagationalgorithm was used to optimize connection weights. The training goal was to produce an output pattern of motor learning (and, , dyskinesia ) that accurately reflected experimentally derived secondarily reported dose responsedata given one of the four levels of diseaseseverity. when the error term (total sum of squared Training was consideredstabilized ' error) between the model s dose responsepredictions and the experimental data was minimized. Model A in figure 14.3 (solid lines) failed to predict treatment responseas based on diseaseseverity, suggesting that improvement in motor response does not directly act through serially connected dopaminergic and striatal units. However, modifications of model A employing direct connections between the striatal unit and either diseaseseverity or levodopa dose units -responsecurves , (dotted lines) satisfactorily modeled the experimental dose

336

in Dementia Applications

suggesting direct mechanismsof action on the striatum. The remaining models were modified in an analogousmanner , yielding additional infonnation on treatment effects , dyskinesia , and potentially significant neuronal connection strengths (Jamieson , 1991).
Motor Maps

In backpropagationnetworks, specific node locations within a given layer have little effect on either input or output values eachnode , largely because within one layer is connectedto every node in the adjoining layers. However " " , for self-organizing networks the spatial distribution of responses within a layer is of primary importance . The organization of intralayer connectivity is optimized to translate similarity of input signals into proximity of activated nodes . This allows for the creation of topographic maps , in which similar input values are translated into proximal spatial relationships among responding neurons in the network layer itself (Kohonen , 1982; Von der Malsburg, 1982; Willshaw 8t von der Malsburg, 1976). The assignment of valuesto lattice sites is modifiable to allow for new learning. Additionally , although the spatial arrangementof outputs within the map is critical, it can be a computationalabstractionof real-world spatial mapping. A critical characteristicof self-organizing networks is the incorporation of lateral inhibition (competitive learning) into the model. This feature was derived from neurophysiological mapping using single-unit recording of , especiallythe visual and somatic systems(Hubel biological sensory systems 8t Wiesel, 1962). In the most basicfeaturemap , a single node is connectedto each node immediately surrounding it. Through learning, connectionsfrom the most active node becomeexcitatory to immediately surrounding nodes , and inhibitory to more remote nodes . Proximity to the active node also determinesthe strength of the connection weights, being inversely related to the distance between the primary (active) node and its neighbors . This createsa dynamic competition between nodes in the layer, with nodes that are remote to the " winning" node either having a weak signal, an inhibitory signal, or no signal at all. Such a system stabilizes so that only the node with the largest overall input value effectively propagates its signal to the " next layer, or becomesan " attractor basin within the feature map itself (Aleksander8t Morton , 1995; Dayhoff, 1990; Freeman8t Skapura , 1992). Self-organizing motor maps represent an extension of sensory feature , allowing for the storage of output values for each matrix point . By maps , a self-organizing feature map can be including learning of output values designed to model motor control paradigms (Ritter 8t Schulten , 1987). In these models , control commandsare mapped to two -dimensional locations of excitement , just as in sensory maps (Lemon , 1988). However, the formation of a map correspondsto the learning of a specificmotor task. In contrast to the flat, two -dimensional feature map of sensory representations , movement occurs in a three-dimensional space . Thus, target aiming and

337

-'s Disease Mahurin: BasalGanglia Function in Parkin ~ r.

manipulation require a three-dimensionalmatrix. This three dimensionalmap does not necessarilyconform to the flat microcolumnar topology of the cerebralcortex itself. Nevertheless , a three-dimensionalgrid can be flattened into a two -dimensionalspacewithout loss of original connectivity and organization of the excitatory nodes necessaryfor algorithmic representation Ritter & Schulten , 1987). ( Figure 14.4 illustrates the basic architectureof a sensorymap and a motor map. The upper portion of the diagram depicts a topographic feature map for recognition of sensoryinput. The secondlayer forms a featuremap based " " on lateral inhibition of competing input. The winning, or active, node is ; proximal nodes are gray; and inhibited nodes are white. A depicted in black of an input vector and feature node. learning step requires correspondence The lower portion of the diagram representsa topographic self-organizing motor map. The addition of the third layer allows for storage of an output . In this case value specific to a motor command , the learning step requires correspondenceof the input vector, feature node, and output vector. The network utilizes selective enhancementof the motor signal through lateral inhibition in order to sharpenprecision of motor control. The biological instantiation of motor maps theoretically corresponds to neuronal populations , rather than individual , such as cortical microcolumns neurons(Goldman Rakic , 1957; Takeuchi& Amari, 1979). , 1984; Mount castle Associationsamong network connection patterns and motor cell assemblies . These associationsinitially are adjusted according to specifictask demands in are establishedthrough trial and error learning the development of motor skills, but later refined through iterations of dosed-loop associativelearning. , basal ganglia activity can be modeled in terms of timeUsing motor maps dependent visuomotor integration, with outputs to cortical and brainstem motor targets. These designs can be expanded to allow for multiple arrays , es (Morita , 1996). Output modeling increasingly complex motor process from motor maps can be scalar , giving , vector, or linear transformations values can . The to the control specify torque, output process great flexibility motor action to force ballistic or sustained linear displacement , allowing , , vary in nonlinear ways depending on the dynamic excitation of the feature layer.
Motor Learning

A self-organizing network also provides a mechanismfor the learning and control of movement. Although biological equivalence is not assumedin these models , the basal ganglia provide a neuroanatomical substrate that could implement a self-organizing feature map. Neurophysiological studies have revealed convergence of inputs from motor and perceptual cortical associationareasinto the more spatially constrained architecture of striatal : Faull, 1989). Within the structures (Alexander et al., 1986; McGeorge & " " a functional to be discrete striatum there appears organization of domains

JJ8

in Dementia Applications

( Gc U0 . ~

I W C . ~ E W . C D

I INPUT VECTOR I

. W, transfer - organizing networks funman ; blacksquares , active Figure 14.4 Self topographic . (Adaptedfrom Ritter 6 : nodes , inhibited nodes , weak nodes ; white squares ; gray - - squares .) Schulten . 1987

ff0

' s Disease Mahurin: BasalGanglia Function in Parkinson

~ ca a

0 C 0 ~ A ~ t > W t ~ tJ

~ ca a

G C 8 w E ~ . C W .

of spiny neurons corresponding to intersectionsof motor and sensory neocortical afferents (Rebec& : Curtis, 1988; Selemon& : Goldman-Rakic , 1985). the medium in cells the striatum exhibit Although spiny overlapping dendritic fields and local axonal aborization neurotransmitter , dependentlateral inhibition has not directly been establishedbetween these elements . Therefore , it has been proposed that passive communication occurs within the striatal matrisome regions by way of direct electrotonic coupling (Connolly & : Bums , 1995). In this manner , the firing frequency of the spiny cells codes , and the gradient information correspondingto both the current feature state next goal state in a motor sequence . Cortical neurons that project to the " basal ganglia set the boundary conditions" for competition within the feature map. Interactions between these striatal regions are hypothesized to be mutually inhibitory , with a spatial gradient function such as that for selforganizing feature maps ( Wickens : Alexander, 1995). This type of , Kotter, & architecture would allow for rapid computation over optimally short communication , which in turn project back to the more sparselycoded pathways cortical regions (Strick, 1985). Maintenance of a network pattern within the striatum (i.e., relative weights of connectivity) potentially is implemented by elevations of spine calciumat corticostriatal synapses , which may persist for time intervals of up . to a second( Wickens 1988 Striatal , ) dopamine is well establishedas fundamental based to contingency learning in animals(e.g ., experimentsinvolving food, water, self stimulation), and hasbeenhypothesizedto mediatea reward mechanismthrough brainstem afferentsto the striatum, potentially reinforcing accurate sensorimotor behaviors (Hirata, Yin , & : Mogenson, 1984). In addition, a high number of corticostriatal inputs terminateon dendritic spines that have convergent dopaminergic input from midbrain nuclei ( Wickens , 1988). Striatal activity has been linked to local dopamine levels which are influencedby both cortical and limbic input (Kalivas : Klitenick , Churchill , & , 1993). Approximately 40 percent of striatal spines are contacted by two boutons, one of cortical input, the other staining positively for dopamine " " -responseassociation : Smith, 1984). Thus, an effective stimulus , Powell, & (Freund (i.e., sensorimotorlearning) can be strengthenedat the synaptic level -dependent convergence of three striatal inputs: cortical sensory time , by cortical motor, and brain stem reinforcement(ie ., concurrent dopamineexcitation , Kelier, et al., 1991; Wickens , 1990; Wilson, 1995). ) (Iriki , Pavlides

COGNITIVE NETWORKS
" " Neurophysiological studiesindicate at least two primary cognitive striato cortical pathways (Alexander & Crutcher, 1990). Both pathways originate in the prefrontal cortex, one in the dorsolateralarea , and the other in the lateral orbitofrontal area . Both pathways have targets in the dorsal caudatenucleus . In contrast to the putamen es , which is primarily involved in motor process ,

340

Applications in Dementia

the caudatenucleusis hypothesized to playa critical role in cognitive functions , including learning, memory, behavioral planning, perception, and , 1992; Cummings , 1990; Mayeux, 1990). A variety of language (Crosson es networks have beenconstructedfor modeling of cognitive process , including those that involve the basalganglia. Many of thesenetworks emphasize cortical elements , and discussion of them can be found elsewhere in this book. One model, that of striatal-mediated spatial attention, is presented below.

SpatialAttention
Attentional dysfunction is a common feature of PO, in which both basal ganglia and cortical regions are closely concernedwith the integration and allocation of spatial attention (Boller, Passafiume , Keefe , et al., 1984; Brown & : Marsden , Reisman , , et al., 1991; Levin, Llabre , 1986; Hutton , Morris , Elias et al., 1991 ; Wright , Bums, Getten, et al., 1990). Frontoparietal connectivity has been describedas underlying a large-scaleattentional network with two : (1) an " anterior" attention system (prefrontal and major cortical components anterior cingulate ) that selects(filters) incoming perceptual information and " " directed action sequences , and (2) a posterior attention system preparesgoal (parietal lobule, pulvinar, superior colliculus) that codes and transmits information regarding the spatial position of behaviorally relevant stimuli : Raichle , 1994). Anterior and posterior attention , 1990; Posner & (Mesulam systems have reciprocal topological homology in their node-to-node connections , implying preciseintegration of spatial information (Goldman-Rakic & : Friedman , 1991). In addinon, both systems synapsein the striatum in an , allowing for integration within this region. alternating columnararchitecture The basic mechanismof this attentional network is comparable to the : Houghton, motor control network depicted in figure 14.2 (Jackson & " " 1995). However, in this instance the input to the comparator is that of -attentional mismatch . The a perceptual , rather than a limb-target mismatch trigger for the basalganglia to propagate a signal is a match mismatchcomparison . This of inputs from the anterior and posterior attention systems in cortex or in the striatum occurs either prefrontal comparisontheoretically itself (Schultz , Apicella, Romo, et al., 1991). When a match occurs (i.e., when the posterior spatial detection of the stimulus matches the anteriordetermined positional goal state of the stimulus), the basal ganglia network is stabilized (i.e., optimized). This increasesinhibition of thalamic output, and continues stabilization of the current system state. If , however, the intended and actual location states differ, the basal ganglia decreaseinhibition of thalamic output, potentiating the flow of thalamofrontal information to initiate corrective motor activity (i.e., to readjust body or limb position " " : Houghton, 1995). This tuning of relative to stimulus location) (Jackson& through the basalganglia is critical to proper spatial attentional mechanisms movement , but may be disrupted in PD.

341

' s Disease Mahurin: BasalGanglia Function in Parkinson

LIMBICNETWORKS
" " The limbic corticostriatal pathway originates in, and returns to , regions of the anterior cingulate and medial orbitofrontal cortex. Subcorticaltargets of this pathway are in the ventral striatum, which has dopaminergic connections with the ventral tegmentum and nucleus accumb ens. As discussed earlier , the cognitive and limbic pathways maintain a dorsal-ventral separation at both cortical and subcortical levels (Helmer, Alheid , & : Alheid, 1993; 1987 see 14.1 . , )( ) Jayaraman figure

-basedLearning Contingency
In conjunction with the limbic system , the basalganglia underlie in the translation of motivationally relevant stimuli to adaptive behavioral responses ( Kalivas et al., 1993 ; Mogenson, Brudzynski , Wu , et al., 1993). The limbic , and monitors the outcome system itself regulatesthe recognition of salient goals of behaviors directed toward achieving those goals. In turn, the basal ganglia carry out stored motor actions to obtain the limbic-generatedgoal states . Since most goals rarely are immediately present , a series of linked must be constructed and executed to subgoals temporally move the organism to where it can obtain eachof a seriesof goal states0 . A . Gray, 1995a , b). There is an extensive literature describing neostriatal mechanisms of behavioral leaning and goal-directed behavior. As with motor networks, severalmodels of contingency-basedlearning hypothesize the formation of learned associationswithin the extended basal ganglia network ( Wickens& : Kotter, 1995). Coding by spiny neuronsof interactions within specificdorsal cell assembliesmay occur in the dorsal striatum (Rolls, 1995). Local connection weights thus are modified according to temporally dependent -mediated feedbackfrom limbic (ventral) dopamine portions of the striatum. toward not Progress directly be monitored by the proximate goals may dorsal striatum, but by limbic structures , including the septohippocampal system and Papez circuit, including a loop from the subiculum via the , anteroventral thalamus : , and cingulate cortex ( Totterdell & mamillary bodies Smith, 1989). In this manner , dopamine projecting neurons from the ventral ens , tegmental area mediate a reward function via the nucleus accumb which interacts with the dorsal striatum in the selection and updating of : Gray, 1992). , 1985; Young, Joseph , & appropriate motor actions (Oades term of based information memory storage contingency Long potentially is maintained in the septohippocampal system , with involvement of other : Granger ; Lynch & , 1992; regions suchas cerebellarnuclei (J. A . Gray, 1995b ' O Keefe & : Nadel, 1978; Rawlins : Seferiades : , Lyford , & , 1991; Schmajuk & Di Carlo, 1991).

342

in Dementia Applications

" circuits, the striatocorticolimbic loop involves another " comparator function . This function continually updatesthe organism regarding the match or ' mismatch between the network s current state and a motivationally salient " " " " goal state. The systemalso monitors the success or failure of eachaction in the goal gradient in approachingeach subgoal . Basedon this information, a backpropagation network dynamically updates the connection weights within the system . Comparator information is updated on a cyclic basis , corresponding " of information to each " pass the limbic basal through ganglia loop (J. A . Gray, 199sb; Weiner, 1991). These contingency-based associations are mappedonto an associativenetwork. This model requirestemporal ' delay in the holding of information within the network, so that the last " " verified state of the world can be held for comparison with the current " state of the world. In this manner the system can gauge progress toward eachsubgoal , as well as toward the final goal state (J. A . Gray, 199sb).
Neuropsychiatric Disorders

A variety of neuropsychiatric disorders are associatedwith basal ganglia , dysregulation of the contingency-based reward dysfunction. For example , system has been associated with motivational deficits in schizophrenia . A . F. Poirier & : 1991 and Scaravilli , , ; , ] Gray, anxiety ( Gray, depression Koob . and : 1995b , 1987) Grossberg ; Richfield , 1993; Swerdlow & colleagues , 1984) have developednetwork modelsfor severalbasalganglia(Grossberg mediatedneurologic and psychiatric disorders , , including PO, schizophrenia These models and attention deficit disorder. , employ hyperactivity depression the circuit-switching device of gated dipoles , an artificial instantiation of . Excellent reviews of neural opponent processingfound in biological systems the gated dipole model can be found in several sources(Grossberg , 1984, 1991; Levine, 1991). Although the conceptsof gated dipoles and opponent processing are discussed less frequently than backpropagation networks, es. they contain powerful featuresfor modeling neuropsychologicalprocess A schematizedgated dipole model, adapted from the work of Grossberg ; see , b) is presentedin figure 14.5. (for further discussion , 1972a (Grossberg " " also Levine, 1991). Synaptic transmitter releaseat synapsesWI and W2 . Phasicactivity undergoesdepletion secondaryto tonic nonspecificarousal WI , and A further increases the of , activity depleting pathway temporarily When this is removed inhibition to the node. , increasing output phasicinput " " a competitive rebound occurs becauseof now greater signal strength in -connected . The cross pathway B, increasingexcitation to the output nodes inhibitory pathways at A3 and B3 augment either effect. When system " " arousal is low , both the on-output and off -output channelsexhibit abnormal . This output is determined by discontinuation of phasic activation properties of one input channel , in conjunction with enhancementof tonic activity from a secondrebound channelonce the input from the first channel is diminished or removed. Increasedactivity of one channel , therefore , is

343

' Mahurin: BasalGanglia Function in Parkinsons Disease

~ ~~ N
I Phasic I I Input I
344

(+)

(+)

Transmitter Release

(+)

(+)
I TonicI I InputI

forward gated dipole network subsystem . I, chronic (nonspecific Figure 14.5 Diagram of a feed ; J, phasic arousal ) arousal ; A, B, network nodes (neurons ); WI , W2, presynaptic nodes ; 0 , output node. (Adapted from Grossberg , 1991.)

consequent to the cessation of phasic activity in the opponent channel. A gated dipole is able to compare current stimulus or reinforcement patterns with recent values of the same variables . The two channels are opposite in " " " " sign and gate signals based on the amount of computed transmitter values released at critical nodes. The gated dipole can model such clinical features as bradykinesia , the " onl " off phenomenon , agitated depression , and manic- depressive mood swings (Grossberg , 1984 ). In these descriptions , the output of the gated dipole impinging on the basal ganglia depends on the interaction between context specific time - limited activity (phasic arousal), and continuous nonspecific activity (tonic arousal). Monoamine levels theoretically determine tonic arousal. If nonspecific arousal is within a normal range, then normal behavior occurs. However , if nonspecific arousal is either above or below this normal range, pathological behavior occurs.

Applications in Dementia

If tonic arousal is too low , the response threshold of phasic inputs is raised; that is, greater context -sensitive phasic arousal is required to activate these connections . Once the input threshold level is achieved , the system becomes hypersensitive to phasic input , consistent with symptoms of distractibility or inattentiveness in pathological states. As a slowly changing tonic mechanism, such switching is the hypothesized situation in attention " " deficit disorder , the on / off phenomenon in PO , and the mood swings " " between mania and depression . The tonic gain can be increased by administration " " to the system of an activating pharmacological agent , such as levodopa (carbidopa ), amphetamine , methylphenidate , or a selective seroto nin reuptake inhibitor (e.g ., fluoxetine ). A secondary effect of this intervention is a lowering of hypersensitivity . In contrast , if tonic arousal is too high , the threshold for response to phasic input is reduced. However , once threshold is achieved , the system becomes relatively insensitive to changes in phasic input . This model of dopamine overactivation (tonic arousal) is consistent with two apparently contradictory symptoms of schizophrenic illness: positive symptoms (hallucinations , delusions , and conceptual disorganization ) and negative symptoms (poverty of speech, motivation , movement , and affect ). Too great an increase of an activating agent such as dopamine may result in schizophrenic -like symptoms (e.g ., amphetamine psychosis ). Overuse of drugs used to treat schizophrenia by lowering dopamine activity may have a contrasting effect , and induce parkinsonian side effects. Another network model of basal ganglia - related psychopathology described by Hestenes and colleagues (Hestenes, 1991 ) attempts to account for interactions among the three primary monoamine neurotransmitters : dopamine (DA ), serotonin (S HT ), and norepinephrine (NE ). An example is manic -depressive illness. In this disorder , the site of dysfunction is ascribed to the nucleus accumbens, whose activity is influenced by both serotonergic and dopaminergic projections from brain stem nuclei . The nucleus accumbens, in turn , modulates motivational input arising from the limbic system : Zahm , 1993 ). This and projecting to striatal targets (Deutch , Bourdelais , & modulation occurs secondary to the relative gain of both DA and S- HT : Koob , 1987 ). brainstem efferents to the nucleus accumbens (Swerdlow & Evidence for this model is derived from several observations . DA pathways from the substantia nigra pars compact a project both to striatal and cortical regions (e.g ., dorsolateral prefrontal cortex ). PO -associated decrease in substantia nigra - produced DA is reflected in decreased receptor binding sites in the striatum , and is associated with decreased initiation and movement speed. Additional DA modulated input arises from limbic structures , which are involved in expression of schizophrenia . In this model , tonic arousal is considered analogous to generalized DA -mediated system activation . NE is considered important for its regulation of attentional processes, particularly selective attention (phasic arousal), as well as limbic modulation of motivation , drive , and reinforcementS - HT is hypothesized to set a level

345

' Mahurin: BasalGanglia Function in Parkinsons Disease

-wide activity that facilitates of system pattern matching between different brain regions through specificand nonspecificinhibitory effects(Swerdlow & : Koob, 1987).
Depression in Parkinson ' s Disease

PD is &equently accompaniedby a persistent depressivestate. It is unclear whether the degree of depressionseen in PD parallels diseaseprogression , suggesting a common pathological mechanism , or if depressionis a reactive responseto the debilitating aspectsof PO, independent of diseaseseverity (Ehmann , Beninger , Gawel, et al., 1990; Gotham, Brown, & : Marsden , 1986; MacMahon & : Fletcher , 1991; Marsden , 1990; Mayeux, Stem, Williams, et al., 1986). We approachedthis problem using neural network modeling of metabolic activity of brain regions involved both in PD and in major depression (Mahurin, Mayberg, Brannan , et al., 1996). Three groups were studied: unipolar major depression(UD ) (n = 25), PD with depression (PDD) (n = 12), and age-matched normal controls ( NC ) (n = 12). The glucose metabolic rate was measuredwith Eluordeoxyglucose positron emission tomography (PET ) . Both depression groups showed similar degreesof depressivesymptomatology (mean Hamilton Rating Scale for Depression= 29.4; SD = 6.9). Twelve regions of interest (ROIs) were included in the model, six previously shown to be affected in depression (dorsal anterior cingulate , ventral anterior cingulate , orbito&ontal cortex, caudatenucleus anterior , temporal region, and inferior parietal lobule) and six that had been shown to be unaffected (superior and inferior temporal gyri , posterior temporal region, hippocampus , thalamus , and striate cortex) ( Mayberg , 1994; Mayberg et al., 1990). Normalized PET counts for each of the 12 ROIs were input into a back network propagation training model, with the three diagnostic groups ) values (figure 14.6). Eight inner representing unique output (classification units were in the network architecture processing , characterized employed summation of data PET counts), a sigmoidal transfer by weighted ( input function, and a 0 to 1 output value. Backpropagationrecurrently adjustedthe connection weights between the network units until the system relaxed into a fixed attractor state with a root mean square classification error of no . greater than 5 percent The network was retested with a randomly preselected10 percent of the original data set, again achieving the samelevel of diagnostic classification for depressed to the groups vs. controls. Addition of two cognitive measures model, the Stroop test and Trail Making test, increasedthe efficiency of both network classifications by a factor of 4. Once trained, the network was able to diagnosis an individual subject with a correct classification rate of 95 percent. These findings support the hypothesis that a common neural substrate involving striatocortical interaction underlies both motor slowing and de-

346

in Dementia Applications

OUTPUT LAYEA (Classification )

E ~ : ~ ~3

INPUT LAYER ( PET Data)

's disease of Parkinson Figure 14.6 Network modelfor classification and majordepression . data values for , basal were derived , from brainneuroimaging Input including with posi ganglia tron emission . 5 PET values A . C inner (PET ) , VO ; , ; tomography layer , unipolar ; PO , depression 's disease Parkinson controls . (Adapted ; NC, normal fromMahurinet al., 1996 .)

pression in PO. An overlap of functional activity in depressionand PO has been well documentedby both clinical and neuroimaging studies (Mayberg et al., 1990). The locus of this deficit may be in the basal ganglia, &ontal , or limbic regions. The described network model is consistent projection areas with clinical observations as to the difficulty in distinguishing " neurological " " deficits in drive and motivation &om " reactive affective changes . This use of classification networks can be extended to more generalized models that relate cerebral , cognitive, and motor impairment in various neuropsychiatric and neurodegenerative disorders(Litran, 1996).

FUTURE DI R Ecn ON S
Although inspired by featuresof biological neurons , it should be emphasized " i.e. that neural networks have varying levels of "biological equivalence ( , ' fidelity to the anatomical systems they purport to model). Weakly speci " fied computational networks can produce results identical to biological systems , but bear little resemblance in architectureto the biological system that " " is modeled . In contrast , strongly specified systemsattempt to be faithful to the structure and connectivity of the biological system , as well as approximating experimentally derived relationships between input and output values . Although both types of networks are modeled on hypothesized neural connedion weights, the latter cannot diredly be measuredbetween , inferencesare made based fundioning neurons in the human brain. Rather on the relationshipbetweeninput strength or pattern and the resulting output (behavior or physiological change ), similar to mapping the firing rate of the visual cortex as a function of visual signal frequency .

347

' Mahurin : Basal Function in Parkinson s Disease Ganglia

As suggestedby Levine (1991), an effective method of modeling largescale neural systems is to use interconnected subnetworks , in which the ) network functions as input for a second output of one (operation-speci6c network, and so on. If connected , the output of the final subnetwork may be fed back into the first (or any of the middle networks), creating a large reentrant . Information flow is dynamic, self- correcting system , and parallel. Further, an important aspect of multiple subnetworks is that they can be modified independently of the need to alter the overall architecture of the . system Figure 14.7 presents a schematic of how interconnected subnetworks might be organized to model basalganglia function. For clarity, certain critical intermediate structures , including pallidum, thalamus , substantia nigra, subthalamic nucleus , and brain stemnuclei, are not shown. Also , many aspects

of interconnected basal subnetworks . Networks : MM , Figure 14.7 Schematic ganglia potential motor map RAM random , backpropagation ; FM, featuremap ; AL associative , ; BP learning ; assodative inhibition ; Cl competitive ; GO , gateddipole ; OA . outstaravalanche ; RF , memory realrrentfeedback function . ; GF , gainfunction ; OF , decay ; T, timedelay

348

in Dementia Applications

or

Dtual FM Cortex

Perce ens DF Sxstem / ~ GF AUGD Accumb Limbic GD / Cortex OA / MM Motor BP , Dorsal Frontal Fr M V

of network architecture (e.g ., cross -connections , convergence , transfer functions have been omitted. The subnetworks include self ) organizing feature , associativelearning for sensorimotor reinforcement mapping for perceptual analysis , dipole gating for contingency learning, and backpropagationfor . Discussion of programming and execution of complex motor sequences these networks can be found in this chapter and elsewherethroughout this -wide reinforcement book. With the addition of time- delayed feedback , and system , very complex models of basal ganglia function and diseaseare . The use of connectedsubnetworksis applicableto other aspectsof possible and canbe employed to model a wide rangeof neuropsychological cognition, . disorders This chapter has focused on basal ganglia networks that provide goaldirected motor control and learning, with an emphasison PD. Schizophrenia and depression were discussedas examples of neuropsychiatric disorders associatedwith basal ganglia dysfunction. Howe~ /er, networks have the , including attention potential to model other basalganglia- related disorders ' deficit hyperactivity disorder, Tourette' s syndrome , Huntington s disease , -compulsive disorder, and basal ganglia toxicity . Neuropsychological obsessive aspectsof these disorders can be explored with the network components discussedabove , including opponent processing , associativelearning, lateral inhibition , and time-dependentfeature analysis . New ways of viewing neuropsychologicalfunction will emergein conjunction with new treatment to thinking in terms of networked strategiesas cliniciansbecomeaccustomed es. From this can arisea fundamentalshift in cognitive operationsand process test construction and interpretation basedon anatomicaland computational models of neural networking.

REFERENCES
Aleksander . New Yark : , I., & Morton , H. (1995 , 2nd ed ). An infrnJucfiml to neural computing International Thomson Press . Computer Alexander . In M. A. Arbib (Ed and .), Thehandbook , G. E. (1995 ). Basal of braintheory ganglia neural networks - 144 . , MA: MIT Press (pp. 139 ). Cambridge Alexander : Crutcher architecture of basalgangliacircuits : , G. E., & , M. D. (1990 ). Functional Neuralsubstrates of parallel . Trends in Neurosdences - 271 . , 13 , 266 processing - thalamocortical Alexander : Delong , G. E., Crutcher , M. D., & ). Basalganglia , M. R (1991 " " " " circuits : Parallel substrates in for motor . Progress , oculomotor , prefrontal and limbic functions Brain Research . 119 145 , 85 ( ), Alexander : Strick of functionally , G. E., Delong . P. L. (1986 ). Parallel , M. R., & segregated organization - 381 circuits . Annual . andcortex Review , 9, 357 of Neurosdence linkingbasal ganglia Beiser " ent : Houk of the basalganglia . Cu , D. G., Hua , S. E., & , J. C. (1997 ). Networkmodels - 190 in Neurobiology . ,7 , 185 Opinions BeneckeR of : Marsden , Rothwell , J. C., Dick . J. P. R , Day , B. L., & , C. D. (1986 ). Performance 's disease - 757 in patients simultaneous movements with Parkinson . Brain . , 109 , 739

349

's Mahurin: BasalGanglia Function in Parkinson Disease

Berardelli . P. R . , Rothwe U : Marsden , A , Dick ,J , J. C., Day of the , B. L., & , C. D. (1986 ). Scaling 's sizeof the first agonistEMG burstduringrapidwrist movements in patients with Parkinson . Journal disease and . , Neurosurgery Psychiatry , 49 , 1273 1279 of Neurology Boller AwneD., Keefe . , Morrow, L., & , F., Passa : Kim , N. C., Rogers ,K , Y. (1984 ). Visuospatial ' s disease in Parkinson : Role of andmotor factors . Archives , impairments of Neurology perceptual - 490 41 . , 485 's disease ., Yeap Borrett : Kwan , D. S , T. H., & . , H. C. (1993 ). Neuralnetworksand Parkinson Canadian Sciences . , 20 , 107 113 Journal of Neurological 's disease . G., & : Marsden Brown , R functionin Parkinson . Brain , C. D. (1986 ). Visuospatial , 109 , - 1002 987 . -space : Burns striatal model . In J. C. Houk , C. I., & , J. B. (1995 : D. ). A state , J. L. Davis ,& Connolly G. Beiser .), Models in the basal . . 163 (Eds ( , of infonnation 177 ) processing gangliapp Cambridge MA: MIT Press . -Vidal -thalamocor ras Contre : Stemlmach . (1995 modelof basal , J. L., & , G. E ). A neural ganglia - 476 ticalrelations in normal andparkinsonian movement . Biological . , 73 , 467 Cybernetics in language andmemory . NewYork Crosson , B. (1992 : GuilfordPress . ). Subcorlicalfundions dementia . New York : Oxford University , J. L. (1990 Press . ). Subcortical Cwnmings : Benson : A clinical , J. L., & . Boston : Butterworth , D. F. (1992 ). Dementia Cwnmings approach Heinemann . : An introduction architectures . New York , J. E. (1990 : Van Nostrand ). Neuralnetwork Dayhoff . Reinhold models of movement disorders of basal . Trends in ). Primate Delong , M. R. (1990 ganglia origin - 285 Neurosciences . , 13 , 281 : Zahm Deutch , A. Y., Bourdelais accumb enscoreand shell : , A J., & , D. S. (1993 ). The nucleus Accumbal andtheir functional attributes . In P. W. Kalivas & : C. D. Barnes .), (Eds compartments Limbic motor circuits andneuropsychiatry : CRCPress . Raton , FL (pp. 45- 88). Boca ., Beninger : Riope Ue . J. (1990 Ehmann , T. S in , R. J., Gawel , M. J., & ,R ). Depressive symptoms 's disease Parkinson : A comparison with disabled controlsubjects . Journal of Geriatric Psychiatry andNeurology , 3, 3- 9. ' s disease Evarts : Caine time in Parkinson . Brain , E. V., Teravainen , H., & , D. B. (1981 ). Reaction , 104 167 186 . , : Graybiel for body representation in the , A. W., & , A. M. (1993 ). Two input systems Flaherty striatal matrix : evidence in the . Neuroscience , 13 , primate monkeyJournal of Experimental squirrel 1120 - 1137 . Freeman : Skapura : Algorithms networks , J. A., & , D. M. (1992 , applications , andprogramming ). Neural -Wesley . New York : Addison . techniques Freund : Smith - immunoreactive , T. F., Powell , J. F., & , A. D. (1984 ). Tyrosinehydroxylase boutonsin synaptic contactwith identified neurons reference to , with particular striatoniagral dendritic . Neuroscience 13 1189 1215 . , , spines -Rakic Goldman . (1984 of prefrontal . Trends in Neurosaences cortex , P. S ). Modularorganization , - 429 7 . , 419 -Rakic Goldman : Friedman , P. S., & , H. R. (1991 ). The drcuitry of workingmemoryrevealed by andmetabolic . In H. S. Levin . Eisenberg : A L. Benton .), Frontal , H. S (Eds anatomy imaging , & lobe anddysfunction . NewYork : Oxford UniversityPress . function

350

Applications

in Dementia

's disease Gotham . G., & : Marsden in Parkinson : A quantitative , A., Brown , R , C. D. (1986 ). Depression - 389 andqualitative . and . , , 49 , 381 Journal analysis of Neurology Neurosurgery Psychiatry ' : Scaravilli s disease andparkinsonian , F., Poirier , J., & . In S. , F. (1991 ). Parkinson Gray syndromes Duckett(Ed .), The human ntn1 OU Ssystem - 199 : Lea& : (pp. 179 of theaging ). Philadelphia pathology . Febiger of consciousness : A neuropsychological . Behavioral , J. A (1995a ). Thecontents Gray conjecture andBrainSdenct S . - 676 , 18 , 659 andbasal : Application , J. A. (1995b to anxietyand ). A modelof the limbicsystem Gray ganglia - 1176 . In MS . . The neurosciences (Ed ), ( , ). Cambridge cognitive SChizophrenia Gazzaniga pp. 1165 MA: MIT Press . in thebasal . In D. Evered LAM . (1984 ). neurochemically Graybie subsystems specified ganglia & : M. O ' Connor(Eds .), Functions : CIBAFoundation 107(pp. 114 of thebasal ganglia Symposium 144 . London : Pitman . ) and avoidance . ). A neuraltheory of punishment , I: Qualitativetheory Grossberg , S. (1972a Mathematical Biosdenct S , 15 , 39- 67. and avoidance . ). A neuraltheory of punishment , II: Quantitative Grossberg , S. (1972b theory - 285 Mathematical Biosdences . , 15 , 253 . (1984 behavioral due to transmitter ). Somenormaland abnormal Grossberg , S syndromes - 1117 . Biological es . , 19 , 1075 Psychiatry gatingof opponent process networkarchitecture for Pavlovian : Reinforcement ). A neural , Grossberg , S. (1991 conditioning attention . In M. L CommonsS . Grossberg : J. E. Staddon .), Neural network , forgetting (Eds , timing ,& models andaction : Erlbaum . , NJ (pp. 69- 122 ). Hillsdale of conditioning Helmer Iheid, G. F., & :A Iheid, G. F. (1993 forebrain : An anatomical , L, A ). Basal organization framework for motor aspects of drive andmotivation . In P. W. Kalivas & : C. D. Barnes .), (Eds Limbic motor circuits andneuropsychiatry Raton : CRCPress . (pp. 1- 43). Boca , FL Hestenes networktheoryof manic - depressive illness . In D. S. Levine & : S. J. , D. (1991 ). A neural - 257 .), Motivation Leven(Eds in neural networks , emotion , andgMI direction , (pp. 209 ). Hillsdale : Erlbaum . NJ . , Yin : Mogenson . (1984 motorcortexto Hirata ,K , C. Y., & , G. J ). Excitatory input from sensory neostriatum and its modification . Brain by conditioningstimulationof the substantia nigra Research , 321 , 1- 8. Hoover : Strick in the basal . SdmCt , J. E., & , P. L (1993 ). Multiple outputchannels , 259 , ganglia . 819 - 821 : Wiesel fields HubeLD. H., & , T. N. (1962 , binocularinteraction ). Receptive , and functional 's visualcortex architecture in the cat . Journal - 154 . , 160 , 106 of Physiology 's distast Huber .J ., & : Cummings : Neurobehavioral . New York : ,S , J. L (1992 ). Parkinson aspeds Oxford University Press . Hutton : Vam1a . (1991 contrast is reduced , J. T., Morris , J. L , Elias , J. W., & ,R ). Spatial sensitivity 's disease in bilateral Parkinson . Neurology - 1202 . , 41 , 1200 -termpotentiation Iriki, A , Pavlides : Asanuma of thalamic , C., Kelier ,A ,& , H. (1991 ). Long input to the motor cortexinduced of tha1amocortical and corticocortical afferents . by coactivation - 1441 . , 65 , 1435 Journal of Neurophysiology ., & : Houghton selection andthe basal : A neural network , G. (1995 JacksonS ). Sensorimotor ganglia model . In J. C. Houk : D. G. Beiser .), Models in . J. L Davis ,& (Eds of information processing - 367 thebasal . , MA: MIT Press (pp. 337 ). Cambridge ganglia

' Mahurin: BasalGanglia Function in Parkinson s Disease

's modelof levodopa . P. W. (1991 in Parkinson Jamieson ). A computational phannacodynamics disease . CliniC R I Neuropharmacology - 513 . , 6, 498 ' . & : D. B. s disease : etiologyand treatment . In SH . Appel , J, Caine , Jankovic (1987 ). Parkinson - 234 - YearBook . ), Current : Mosby . (Ed (pp. 193 ). St. Louis Neurology andcognition : An interpretation of anatomical , A. (1987 andconnectivity ) . Thebasal Jayararnan ganglia . In J. S . Schneider & : T. I. Udsky (Eds .), Basal and : behavior patterns ganglia Sensory - 160 : HansHuber . (pp. 149 ). Bern of motor aspects functioning Kalivas : Klitenick , P. W., Churmill the translation , L., & . M. A. (1993 ). The circuitrymediating of motivational stimuliinto adaptive motor responses . In P. W. Kalivas & : C. D. Barnes .), (Eds - 287 Limbic motor drcuits andneuropsychiatry : CRCPress . (pp. 237 Raton . FL ). Boca - organized formationof topologically correctfeature Kohonen . T. (1982 . Biological ). Self maps Hcs , 43 , 59- 69. Cyberne Kotter : Wickens . (1995 of glutamate and dopamine in a computational , R., & ,J ). Interactions - 214 modelof the striatum . Journal Neurosciences . , 2, 195 of Computer Lemon . (1988 motorcortex : Trends in Neurosciences ,R ). Theoutputmapof theprimate , 11 , 501 . 506 -Ramos Levin : Brown , B. E., Uabre , M. M., Reisman , S., Weiner , W. J., Sanchez , J., Singer , C., & , ' s disease M. C. (1991 in Parkinson . 41 . 365 369 ). Visuospatial , , Neurology impairment Levine . (1991 to neural andcogni Hve . Hinsdale : Erlbaum . , D. S ). Introduction , NJ modeling Utvan . E., Jeninger , I., Deleo, J. M., Hauw , J. J., DanielS , K., McKee , A , Dickson . D., Horoupian , D. S ., Lantos : Tabaton teachus about , P. L., & , M. (1996 ). What canartificialneuralnetworks - 839 disorders with extrapyramidal features ? Brain . , 119(Pf3), 831 neurodegenerative : Granger , G., & , R. (1992 ). Variationsin synapticplasticityand types of memoryin Lynch -hippocampal cortico networks . Journal Hve Neuroscience - 199 . , 4, 189 of Cogni ' : Fletmer in Parkinson s disease . Journal MacMahon . D. G., & , P. (1991 ). Depression , of Neurology andPsychiatry . , 54 , 666 Neurosurgery Mahurin : Pirozzolo , R. K., Feher , E. P., Nance , M. L., Levy , J. K., & , F. J. (1993 ). Cognitionin 's disease Parkinson andrelated disorders . In R. Parks . Wilson(Eds :R .), Neuropsychology , R. lec , & 's disease andother dtmtn Has - 349 . Press (pp. 308 ). NewYork: Oxford University of Alzheimer Mahurin ., & : Fox , R. K., Mayberg , S. K., Brickman , J. S , P. T. (1996 ). Diagnostic , H. S., Brannan and assessment in : A neural network utility of PETmetabolic mapping cognitive depression model . Neuroimage . , 3(Pf2), S78 motorfunction of thebasal : TheRobert Warten Marsden , C. D. (1982 ). Themysterious ganglia - 539 . Neurology . , 32 , 514 berglecture ? In E. Bock Marsden , C. D. (1987 ) . What do the basalgangliatell premotorcorticalareas . - 295 M. O. O'Connor . Marsh(Eds .), Motorareas :J cortu (pp. 282 : ,& ). New York of thecerebral Wiley. 's disease - 952 . Lancet . Marsden , C. D. (1990 , 335 , 948 ). Parkinson studies of depression in neurological disease . In SE . ). Neuroimaging Mayberg , H. S. (1994 Starkstein & : R. G. Robinson .), Depression in neurologic disease . Baltimore : JohnsHopkins (Eds . Press University . E., Sadzot ,S , B., Preziosi , T., Andrezejewski , P. L., Dannals , R. F., Mayberg , H. S., Starkstein : Robinson . G. (1990 in theinferiorfrontallobe , ,R ). Selective WagnerH. N. J., & hypometabolism ' s disease in depressed with Parkinson . Annals , 28 , 51- 64. of Neurology patients

352

Applications in Dementia

's disease . Neuro : A reviewof cognitiveand psychiatric disorders , R. (1990 ). Parkinson Mayeux R I Neurology . Rvior R , 3, 3- 14 , tmdBeh Iry, Neuropsychology psychi andParkin . B. W., Sano : Cote , L. (1986 ,J , M., & ). Depression , R., Stem , Y., Williams Mayeux ' .), Advances in Neurology : K. J . Bergmann sons disease . In MD . Yahr& (pp. 451 455 ). New (Eds York: Raven Press . from the cerebral of the projections : Faull , A. J., & , R. L. M. (1989 ). Theorganization McGeorge - 537 . in the rat. Neuroscience cortexto the striatum , 29 , 503 -scale for networksand distributed Mesulam , M. M. (1990 ). Large neurocognitive processing . . Annals - 613 attention , 28 , 597 , andmemory , language of Neurology . In M. A. Arbib (Ed .), The handbook andtheoretical . C. (1995 Miall, R , biological ). Motor control - 600 . networks andneural , MA: MIT Press (pp. 597 ). Cambridge of braintheory : Yim, C. C. Y. (1993 ). Frommotivation , S. M., Wu, M., Yang , G. J., Brudzynski , C. R., & Mogenson -ventral nucleus accumb ens of limbic : A review of dopaminergic to action regulation In P. W. in . involved limbic motor nucleus circuitries um integration pallid pedunculopontine - 236 and neuropsychiatry .), Limbicmotorcircuits Kalivas& : C. D. Barnes ). Boca (pp. 193 (Eds . Raton : CRCPress , FL . mechanism of sequential Morita , M. (1996 ). Computational memory studyon theneural pattern - 146 . BrainResearch andCognitive BrainResearch , 5, 137 's of cat of single neurons Mount castle , V. B. (1957 ). Modality and topographic properties . cortex . Journal somatic , 408 434 , 20 of Neurophysiology sensory 's disease in Parkinson : : Fabbrinii , J. L., & , G. (1988 Mouradian , M. M., Juncos ). Motor fluctuations . . PartII. Annals mechanisms Central , 24 , 372 378 of Neurology pathophysiological . movements Mushiake : Strick , H., & , P. L. (1993 ). Activity of pallidalneurons duringsequential . Abstracts , 19 , 1584 Society for Neurosciences in theCNS . between Oades , R. D. (1985 ). Theroleof NA in tuningandDA in switching signals - 282 . andBiobehavioral Review Neuroscience , 9, 261 . darendon asa cognitive O'Keefe : Nadel , U.K.: Oxford , L. (1978 ). Thehippo , J., & map campus . Press University in primate of information Parent : Hazrati , L. N. (1993 ). Anatomical , A, & processing aspects - 116 in Neurosciences . . Trends basal , 16 , 111 ganglia of basal of the functional : Young , G. R., & ). Speculations anatomy ganglia Penney , A. B. (1983 Neuroscience 6 73 94. Review . Annual disorders , , of . . NewYork: Scient i AcAmerican : Raichle Posner , M. E. (1994 , M. I., & ). Images Ubrary of mind to developnonspatial : Seferiades Rawlins , A. (1991 , G., & ). Doesit stiUmakesense , J. N. P., Lyford - 286 . ? Hippo function of thehippocampal theories ,1 , 283 campus and inhibitionin the neo zones of excitation : CurtisS. D. (1988 Rebec , G. V., & ). Reciprocal - 635. striatum . Synapse , 2, 633 's disease disorders andothermovement motorcircuitin Parkinson . (1993 Rich Aeld , E. K ). Mesolimbic andneuropsychilltry circuits . Limbic motor W. Kalivas & : D. Barnes . In P. C. (pp. 359 (Eds ), . : CRCPress 372 Raton , FL ). Boca ' s self - organizing Kohonen . (1987 Ritter : Schulten , K , H., & ). Extending algorithmto mapping .), Neuralcomputers : E. von der Malsburg(Eds . In R. Eckmiller& learn ballisticmovements - 406 -Verlag . : Springer ). Heidelberg (pp. 393 , Germany

353

' Mahurin: BasalGanglia Function in Parkinsons Disease

Rolls . T. (1995 andconsciousness ,E ). A theoryof emotion to understanding , andits application - 1106 theneural basis of emotion . In MS . Gazzaniga .), The neurosciences (Ed (pp. 1091 ). cognitive MA: MIT Press . , Cambridge Rolls : Williams andmovement , E. T., & , G. V. (1987 related neuronal ). Sensory activity indifferent of the primatestriatum . In J. S. Sdmeider & : T. I. Kidsky(Eds .), Basal and regions ganglia - 59 behavior : Sensory andmotor : Huber . (pp. 37 ). Bern aspects functioning Schell ., & : Strick , G. R . P. L. (1984 ). The origin of thalamic and inputsto the arcuate premotor motorareas . Journal - 560 . , 2, 539 of Neuroscience supplementary : DiCarlo , N. A., & of hippocampal , J. J. (1991 modulation of classical ). Neural Schmajuk dynamics . In M. L. Commons S. & : . E. Staddon .), Neural network , models J (Eds conditioning Grossberg , of - 180 andaction : Erlbaum . (pp. 149 , NJ ). Hillsdale conditioning ., & : Scamati Schultz , W., Apicella , P., Romo - dependent ,R , E. (1991 ). Context activity in primate striatum events . In J. C. Houk : D. G. Beiser , J. L. Davis ,& reflecting pastandfuturebehavioral - 27 .), Models in thebasal (Eds : MIT Press . ofinformation (pp. 12 , MA ). Cambridge processing ganglia -Rakic Selemon : Goldman , L. D., & corticalandsubcortical , P. S. (1985 of ). Common targetareas the dorsolateral and in cortices the rhesus . Neuroscience prefrontal posterior parietal monkeySociety of Abstracts . , 11 , 323 Smith : Bolam , A. D., & , P. (1990 ). The neuralnetworkof the basal gangliaas revealed by the - 265 connections of identified neuron es . Trends in Neurosciences . , 13 , 259 studyof synaptic 's disease : Mayeux in Parkinson Stern , Y., & . In MD , R. (1986 . Yahr& : ). Intellectual 's disease . impairment - 408 K . J. Bergmann .), Parkinson : Raven . Press (Eds (pp 405 ). NewYork Strick . P. (1985 to the corticalmotor ). How do the basalgangliaand cerebellum gain access - 124 areas ? Behavioural Brain Research . , 18 , 107 " in the basal : Mushiake Strick . P. L., Hoover for "outputchannels , J. E., & , H. (1993 ). Evidence andcerebellum . In N. Mano . Delong (Eds : M. R .), Role , I. Hamada and ,& ganglia of thecerebellum - 180 basal in voluntary movement : Elsevier . (pp. 171 ). Amsterdam ganglia Swerdlow ., & : Koob , N. R anddepression , G. F. (1987 : Toward , schizophrenia ). Dopamine , mania a unifiedhypothesis of cortico - striato - pallidothalmic function . Behavioral andBrainSciences , 10 , - 245 197 . Takeuchi : Amari of topographic , A., & andcolumnar , S. (1979 . microstructures ). Formation maps . , 35 , 63- 72 Biological Cybernetics 'ev S. V. Tolkunov : Afanas , B. F., Orlov, A. A , & , (1997 ) . Model of the functionof the neo striatum andits neuronal behavior in . Neuroscience Behavior and P I , , activityduring monkeys Iysi . , 27 , 68- 74 ology Totter dell : Smith of hippocampal , S., & anddopaminergic , A D. (1989 ). Convergence inputonto identified neurons in thenucleus accumb ensof the rat. Journal andNeur , 2, of Chemistry Mnatomy - 298 285 . Von derMalsburg of iso -orientation domains ). Outlineof a theoryfor the onto , C. (1982 genesis in thevisualcortex . Biological , 45 , 49- 56. Cybernetics Weiner ens - substantia , I. (1991 . ). The accumb , mismatchand amphetamine nigra pathway - 55. Belravior R I tmdBrRin Scitna S , 14 , 54 Wickens but chemically isolated synapses , J. (1988 : Dendritic sPinesand ). Electrically coupled - 528 calcium in a rolefor synaptic . Progress modification in Neurobiology . , 31 , 507 -mediated Wickens in motoractivation andreward , J. (1990 . Journal ). Striatal dopamine learning
, 80, 9- 31. of Neural Transmissions

354

in Dementia Applications

Wickens : Kotter . (1995 models of reinforcement , J., 6 . In J. C. Houk ,R . L Davis ). Cellular : D. ,J ,6 G. Beiser .), Models in thebllSal Ii4l(pp. 188 (Eds - 214 of in /Omultion ). Cambridge , processing gRng MA: MIT Press . Wickens .R . , Kotter ., 6 : Alexander ,J ,R of localconnectivity , M. E. (1995 on striatal ). Effects function : stimulation andanalysis of a model .5 20 281 . 298 , , ,vnapse Willshaw : von derMalsburg , D. J., 6 neural connections canbe setup ). How patterned , C. (1976 be self - organization . Proceedings Rl Society B: Biological Sciences , Series of theRoy , 194 of London , - 445 431 . of corticalneurons Wilson , C. J. (1995 to the firing patternof striatalspiny ). The contribution neurons . In J. C. Houk : D. G. Beiser .), Models , J. L Davis Ationprocessing ,6 in the (Eds of in / orm bllSal R : MIT Press . (pp. 30- 50). Cambridge , MA gRngli . J ., Geffen : Geffen , M. J., Bums ,R , G. M., 6 of visual , LB . (1990 ). Covertorientation Wright 's disease attentionin Parkinson : An impainnent in the maintenance of attention . Neuropsycho - 159 R . , 28 , 151 log J 'C : Gray , M. H., 6 release in vivo in , J. A (1992 ). Increased Young , A M. J., Joseph dopamine nucleus accumb ens andcaudate nucleus of therat duringdrinking : A miaodialysis . Neuro study - 876 scimct . , 48 , 871

355

's Mahurin: BasalGanglia Function in Parkinson Disease

15

Neural Network Modeling of Wisconsin Card Sorting and Verbal Fluency Tests: Frontal Lobe- Damaged Applications with ' and Alzheimer s DiseasePatients W. Parks andDanielS . Levine Randolph

THEORETICAL MODELING OF FRONTAL LOBE FUNCTION Several authors have developed neural network models of the effects of frontal lobe lesions on the performanceof particular cognitive tasks (Bapi & Levine , 1994 , 1997; Carter, Mintun , & Cohen, 1995; J. D. Cohen , Dunbar, -Schreiber McClelland, 1990; J. D. Cohen & Servan 1992 Dehaene & , ; 1989 1991 & Farah 1993 Leven & Levine, 1987; , , ; Kimberg , ; Changeux Levine & Prueitt, 1989 ; Monchi & Taylor , 1995; Parks & Cardoso , 1997; Parks , Levine, Long, et al., 1992). All these are models of selectedclasses of or behavioral data. are based on of neuro They cognitive approximations and models of the functions of brain areas(Parks anatomy , general specific Long , Levine, et al., 1991). The data modeled illustrate general themesabout . The modules and components significant deficits that occur with frontal lesions of thesenetworks typically illustrate general organizing principles that were designed to reproduce classesof cognitive data , and are based on structures that may be found in the brain. The reader is cautioned that the " models emerging from our methodology are neither strict '1ocalization " " models . The , nor are they representativeof an equipotentiality approach methods used in this chapter focus on the proportional contributions from multiple brain regions that work in concert to form cognitive networks or , 1997; Luria, 1980; Parks , Crockett, & McGeer, 1989; Parks , systems(Fuster Crockett, Tuokko, et al., 1989). In this chapter , we focus on two tests commonly used by clinical neuropsychologists : the Wisconsin Card Sorting test ( WCSn and the verbal fluency test (Spreen& Strauss , 1998). Both test the ability to follow dynamically instructions and to generate spontaneous , yet planned , behavior changing Kolb & Whishaw . 1995 Performance on both tests is , ( ) particularly impaired by damage to the dorsolateral prefrontal cortex (DLPFC ); in the case of verbal fluency, it is especially left DLPFC lesions that impair performance , Talley, et al., 1993; Janowsky (Heaton, Chelune , Shimamura , Kritchevsky, et al., 1989; Mangeis, Gershberg & 1996 Perret 1974 , ; , ; Robinson , Knight, Heaton, Lehman et al. . have suggested , , 1980) However, other researchers

TEMPLATE CARDS

INPUT CARD

... ... ...

RED

GREEN
YELLOW

BLUE

15 .1 Visual of theWisconsin testused in clinical Card . Figure Representation Sorting practice that the WCST and verbal fluency tests have only weak sensitivity to &ontal lobe damage(for review, seeReitan & Wolfson, 1994). In the WCST, the subjectis given a sequence of 128 cards , eachdisplaying a number , a color, and a shape , and askedto match eachcard to one of four , three yellow crosses , four template cards (one red triangle, two green stars blue circles ; see figure 15.1). The clinician then says whether the match is , the right or wrong, without saying why . After ten correct color matches clinician switches the criterion to shape , without warning. After ten correct matches the criterion is switched to number, then back to color, and , shape so on. Milner (1963, 1964) showed that most patients with damage to the dorsolateral pre&ontal cortex can learn the color criterion as rapidly as normals . , but then cannot switch to shape There are several versions of verbal fluency tests used in clinical practice , 1996; Spreen& Strauss , 1998). Two of the most &equently used versions (Lezak include generating words in responseto a letter or category in one minute. Three letters of the alphabet are used to obtain a total score over a three-minute period. The letters F,A ,S is one version (Benton, 1968). An alternate form using the letters C,FiL or P,R, W is useful for retest assessments , & Sivan , 1994). Additional rules include not allowing (Benton, Hamsher names or proper repeating the same word with a slightly different ending. Our network is primarily concernedwith phonemic or letter fluency using the letters C,FiL. This was done to avoid confusion with categorical , vegetables , and cities), which potentially fluency tests (e.g ., namesof animals have different regional cerebral network associations ( Moscovitch , 1994; Troyer, Moscovitch, & Winocur, 1997).

358

Applications in Dementia

The goal of our modeling , like that of some of the other authors in this book, is to createa network which, for different values of the network variables , can simulate both normal and frontally damaged behavior on these tests. Hence es are involved in , we need to analyze which cognitive process normal test performance . One characteristicset of defectscausedby DLPFC lesions on both tests involves perseverationin a categorization that was formerly , but is no longer, rewarding or correct. Therefore , our networks need to include models of the processof categorization , selective attention, and formation of cognitive habits. We take the approachof reducing theseprocess es still further into smaller ones . These smaller process es , in different combinations , model a much wider range of cognitive and behavioral data, covering areas such as conditioning , decision making, and perception , as well as neuropsychological . The component process tasks es are modeled by principles that have been extensively studied mathematicallyand computationally. This approachwas , 1992, 1998; Levine, 1991; pioneered by Grossberg (1982, 1988; Hestenes seealso chapter4 in this book).

NEURAL NETWORK PRINCIPL ~


Eachof the principles that we commonly use is designedfor performanceof a specificsubfunction . Theseprinciples include 1. associative , to enable strengthening or weakening of connections learning betweenevents basedon contiguity or probable causality ; 2. lateral inhibition , to enable choices between competing percepts , drives, or behaviors , , plans ; categorizations 3. opponent to enable selective enhancement of events that change , processing over time; 4. neuromodulation , to enablecontextual refinementof attention; and S. interlevelresonant , to enable reality testing of tentative classi feedback fications. Most of these principles of network architecture were originally suggested by behavioral and cognitive data. These data were modeled using particular network dynamics that were analyzed mathematically . The mathematics often reflects generic relationships that can be manifested by an in-depth study of the experimental literature and theoretical conjecturing (Levine, 1991; Levine, Parks : Prueitt, 1993; Rolls & : Treves ,& , 1998). The conjecturing has focused the theory and accountsfor the specificationof the neural network principles as follows: . Active representation . Information is representedin a module by stableactivity . patterns . Competitive selection . Ambiguities and inconsistencies among module inputs are resolved by competition.

359

Parks& : Levine: Wisconsin Card Sorting and Verbal Fluency

. Associativelearning . Information is stored in modifiable synapses that encodedcorrelation betweenpre- and postsynapticactivities . Opponent for rapid switches in pattern selection processing . Resonant . New information is compressedand encoded codingand recognition . Recognition is a resonantmatching process . by adaptive resonances

Often, neurobiological or neurochemicalresearchwill suggest improvements in the formulation of generalprinciples that are descriptive of underlying in a psychological , associativelearning, discussed ontology . For example context by Pavlov (1927) and Hull (1943), is thought to be basedneurally on plasticity or modifiability of synaptic strengths . Synaptic plasticity was Hebb 1949 to Pavlovian ( ) proposed by explain conditioning data. Hebb that such was based on physical growth of synaptic proposed plasticity knobs. Subsequentexperimental demonstrations of plasticity have been , starting with Kandel and Tauc (1965), working in the sea slug, widespread and Bliss and Lemo (1973), working in the rabbit hippocampus . , Aplysia Theseexperimenters however did not find actual of , , growth synaptic knobs, but rather some more subtle changesin their electrochemicaltransmission properties. The other four principles likewise have undergone evolution as more experimentaldata have becomeavailable . Lateral inhibition was first observed by psychologists studying vision : Ratliff, 1957) who noted that bright spots were seenas brighter (Hartline & if closer to dark areas , and explained this in terms of inhibition between retinal or cortical reports for neighboring visual field positions. Gradually, this was generalizedto other process es besidesvision, suchas categorization , motor control, and concept formation. In all these cases , models have been built on competitive inhibition between cell populations that code different stimuli or concepts of the same type (Amari, 1977 ; Grossberg , 1973; Von der Malsburg, 1973). Contrast enhancementresults if the lateral inhibition " " is moderate , and selective attention to one component ( winner-take-all ) results if the lateral inhibition is strong. Opponent processing was developed by psychologists to explain shifts from one affect to the opposite affect with a changein the stimulus producing the emotional response(Solomon & : Corbit , 1974). Grossberg (1972a ,b) applied it to a network architecturewhich could processeither novel stimuli or changesin external reinforcement . Grossbergsuggesteda biological basis for opponent processingthat involved transmitter depletion at synapses , but a more plausible basis likely involves allosteric receptors that have been found more recently at postsynapticmembranes , 1981). (Changeux Neuromodulation is a concept with a variety of meanings , but perhapsthe best is given by Kaczmarekand Levitan (1987). These authors define neuromodulation " as the ability of neurons to alter their electrical properties in responseto intracellular biochemicalchangesresulting from synaptic or hormonal stimulation" (p. 3). In networks, the idea of modulation is suggested

360

in Dementia Applications

by a large amount of cognitive data whereby contextual shifts influence what sensory features or stimuli are attended to by the organism . These contextual shifts could arise either externally, from changesin reinforcement , or internally, from changes in drive levels. Such attention contingencies shifts have suggestedto many neural modelers(Dehaene& : Changeux , 1989; Leven & : Levine, 1987) the possibility that one neural signal can biochemically modulate the strength of another signal along a different pathway. This 's notion is instantiatedboth by the synaptictriads of Dehaeneand Changeux ' model, and the biasnodes of Leven and Levine s model; the latter is discussed below under Wisconsin Card Sorting Model. Modulation is also suggestedby a range of neurochemicaldata about the monoamine transmitters (norepinephrine , dopamine , and serotonin) and . Eachof these substances is produced and releasedby adifferent acetylcholine midbrain nucleus , then sent out to broad areas of the cerebral cortex and limbic system . There is enough fragmentary evidence (Hestenes , 1992, 1998) to suggest network roles for all three of these transmitters (for , selective enhancementof novel or significant inputs; for norepinephrine serotonin, testing of match between two patterns ; for dopamine , potentiation of reward signals . ) Interlevel resonant feedbackwas introduced into a computational architecture model of by Carpenterand Grossberg(1987). The adaptiveresonance and which was one of the bases for our WCST model, Grossberg , Carpenter was discussedin chapter 1. Basically it includes two of network , layers nodes , a featureand a category layer, which may correspondto two different levels of sensory processing . Arrays of feature patterns are classified into categories , and comparedwith previously stored prototypes (basedon weighted averagesof previous sensory inputs) to seeif they are sufficiently close to the prototype patterns . If a new input pattern arrives that is unlike any prototype , it is assignedto a new category node. Variations of these principles are introduced again , as needed , in the next two sectionsin the context of modeling the WCST and verbal fluency. These models include deficits on the WCST with dorsolateral prefrontal , and verbal fluency performancewith prefrontal damageor different damage ' . of stages Alzheimer s disease

THEWISCONSIN CARDSO RnNG MODEL


Leven and Levine ( 1987 ) simulated the card sorting data using the network of figure 15.2. In this network , based on adaptive resonance theory (ART ) (cf. figure 1.5 in chapter 1), the nodes in the layer FI code features, whereas the nodes in the F2 code template cards determine categories . FI divides naturally into three parts that code colors , numbers , and shapes. (How these are abstracted out from the raw pattern is not considered in this network . There are numerous models of perceptual segmentation in the neural

361

: Wisconsin Parks & : Levine CardSortingandVerbalFluency

CATEGORIES (F ~
BIASES
REINFORCEMENT
~ ' . : ' I I _ _ _' _ II _' I I ) : : : : : ~ : : : (

+ or . : ' " ~

ATTENTIONAL

GATING
0

HABITS
. Neural network model of the Wisconsin Card Sorting test used to representdif Figure 15.1 and nonnal control subjects . ferences between Erontallobe patients

network literature, such as Grossberg and Mingolla , 1985.) Corresponding " " " to each classof feature (color, number , shape ) is a habit node and a bias node." The habit nodes code how often classificationshave been made , feature . The bias nodes or on the basis of each additively wrongly , rightly ' combine habit node activities with reinforcementsignals (the experimenters " " " " right or wrong ). The bias nodes in turn modulate the excitatory signals from Fl to Fl , The network of figure 15.2 utilizes three of the network principles discussed in the last section . Lateral inhibition is used at the layer Fl ' to allow decision between competing categorizations of an input card. interlevel resonant feedbackoccurs between the two layers Fl and Fl , This does not, however, involve long-term associative(Hebbian) learning, as in the classic : Grossberg , 1987), for the connectionsare applicationsof ART (Carpenter & . This short-term assumedonly to be modified by the duration of the task modification of weights is assumedto be handled by neuromodulation via . Pennington (1994) has also suggestedthat connectionsfrom the bias nodes there are connectionsto some parts of the frontal lobes that are not modifi. able , in order to facilitate transient working memory storage

362

Applications in Dementia

QR

SHAPE

MJMER

FEATURES (F1 )

MATCH SIGN 0 0 0 0 M00000SHAPB & 0 1U -

Table 15.1 Resultsof Erontallobe modeling using the Wisconsin Card Sorting test with Eron Tailobe patients and control subjectcomparisons Criterion IX= 4 ( Normal ) Color Shape Number Color (again) IX= 1.5 (Frontally damaged ) Shape(again) Color Thereafter , classifiedby color for all remaining trials Trial 13 40 82 96 115 13

Table 15.1 shows the results of simulating this network. A network parameter measuring the gain of reinforcement signals to bias nodes was varied. The network with high gain acted like Milner ' s normal subjects , whereasthe network with low gain acted like Milner' s frontal patients. The interpretation of prefrontal damage as weak reward signals is supported by the findings of Nauta (1971), that the prefrontal cortex is the area of neocortex with strongest connectionsto motivational areasof the hypothalamus and limbic system . The separation of memories and habits is supported by results such as those summarizedin Mishkin , Malamut, and Ba chevalier (1984). These researchers showed that one neural system subserves memories of the reinforcement value of stimuli, and another system subservesmotor habits regardless of reinforcement . The memory system includes the hippocampusand amygdala (part of the limbic system ), and the habit system includes parts of the corpus striatum. Moreover, evidence reviewed in N. ] . Cohen and Eichenbaum(1993) shows that hippocampally with deficits in both semanticand episodic memory are damagedamnesiacs far lessimpaired on implicit memory for specificmotor procedures . Besidesthe model of WCST introduced in Leven and Levine (1987) and Levine and Prueitt (1989), the other major neural network model of this task was developed by Dehaeneand Changeux (1991). As discussedin Levine, Parks , and Prueitt (1993), there are striking parallels between the two models , even though they were developed independently : one primarily from the neuralnetwork principles sectionof this chapter , the other (Dehaene from ) considerations . Changeux primarily neurophysiological -Changeux model. The Figure 15.3 shows the Dehaene analogiesbetween " " the input" and " memory" nodes of figure 15.3 and the " feature and " category " " " fields of figure 15.2, and between their rule-coding clusters and our " " bias nodes . As in our model, their rule clustersmodulate , are extremely close from input to memory nodes . Their equationsdemonstratethat synapses their rule-coding clustersbehavesimilarly to ours. In both models , the activity of rule clusters is an analog variable that can increaseor decreasefrom

363

Parks & : Levine : Wisconsin CardSortingandVerbalFluency

Reward Error cluster Rule- coding clusters Autoevaluation loop Memory

Current intention

Input

Output

-Changeux test . 15.3 Dehaene neural network model of theWisconsin Card Sorting Figure one card presentation to the next. Our model also includes explicit lateral inhibition between rule clustersto enhancecontrast , which theirs does not. ' Dehaeneand Changeuxs model of the WCST goes through six stagesof . complexity, characterizedby how they choose rules to make decisions These choice strategies are described , in order, as (1) random; (2) random ; (4) reasoning , no memory; (5) reasoning plus context; (3) random plus memory " " plus memory; (6) optimal. Memory refers to the episodic memory ) mentioned above. capacity (via an opponent-processing like mechanism " " all rules that would lead to the sameincorrect refers to Reasoning rejecting have been reached the most as would by recently rejectedrule. The response last two capabilities , reasoning and episodic memory, were not included in the WCST model of Leven and Levine (1987), mainly becausethe authors thought that the WCST was not sensitive enough to require such capabilities . However, Levine and Prueitl (1989) also use a similar network to model data (Pribram , 1973) showing enhanced reaction to novelty in monkeys with damage to the orbital prefrontal cortex. Levine and Prueitl included in the " novelty " model an opponent-processingmechanismwhich ' behavessimilarly to Dehaeneand Changeuxs episodic memory.

364

Applications

in Dementia

DORSAL VS . ORBITAL PREFRONTAL CORTEX AND EXECUTIVE FUNcn ON


The network of figure 15.2 does not deal with the distinctions between different parts of the prefrontal cortex. Petrides (1994) discussedrelevant dissociationsbetweendorsolateral(DLPFC ) and medial and orbital (MOPFC) prefrontal cortex. After summarizingselectedmonkey and human lesion data on different cognitive tasks , Petrides suggestedthat tasks beyond a certain such as those where the concept of novelty is manipulated , complexity , seem to involve MOPFC. Petrides stated that tasks of still greater complexity, suchas those that entail deductive reasoning , seemto involve the DLPFC. The data, however, do not unequivocally support the notion that the main differencebetween those two regions is one of complexity. For example , the task which is , delayed response , is impaired primarily by relatively simple DLPFC lesions . From other data describedby Petrides , the main reason for DLPFC involvement in delayed responseseemsto be the spatial nature of the task. That is, since the two objects are identical, in order to distinguish them the monkey must store their locations. Other tasks that are nonspatial -matching-to are more dependenton MOPFC. In fact, on the delayed sample task , there seemsto be a parcellation whereby dorsal cortex is involved if , the importance of spatial cuesare helpful, and orbital otherwise. Presumably -related processingis related to the connectionsof this the DLPFC in spatial -sensitive regions of the region with space parietal cortex, whereasMOPFC modulates nonspatial processing through connections with object-sensitive regions of the inferior temporal cortex. Another distinction between DLPFC and MOPFC is that the former is more involved in detailed cognitive processing , whereasthe latter (via reciprocal connectionswith the amygdalaand hypothalamus ) is more involved in affective processing . In both the cognitive and affective subsystems , inhibition of inappropriate responsesis important. Milner and Petrides (1984) noted that DLPFC is the most critical area for WCST, whose performance relies on inhibition of habitual responseswhen the context makes these . Yet Damasio (1994) reported findings on patients responsesinappropriate with MOPFC lesions suffering lack of inhibition in socially inappropriate situations . Recentresults of Dias, Robbins , and Roberts (1996) support such a functional dissociation of attentional and affective shifts at the prefrontal level. Current efforts are underway (Levine, 1996) to integrate both the dorsal and orbital areasof the frontal lobes , and their connectionsto basalganglia, thalamus and , amygdala , , into an overall neural network theory hippocampus of what has widely becomeknown as executive , 1996; Luria, function (Lezak 1980). The term has been given many definitions, among them, " to maintain an appropriateproblem-solving set for attainment of a future goal" ( Welsh& : " , 1988, p. 201), and to select the responsesmost advantageous Pennington for the organism in a complex social environment" (Damasio , 1991, p. 404).

365

Parks& 1 : Levine: Wisconsin Card Sorting and Verbal Fluency

Other researchershave suggested that executive functions also include insight, self-control, self-monitoring, and self planning (Karpov & Haywood, 1998; McGlynn & Schacter , 1997; Rabbitt, 1998). " " Many neuropsychologists have avoided the word executive for the frontal lobes becauseof its connotation of a single brain area with highly " " , precise functions controlling the functions of the rest of the brain (Stuss Shallice , & Alexander, 1995). However, we believe that the executive concept is a useful one for understanding the functional integration of many interdependentbrain areasinvolved in cognitive function in complex environments but of . It is particularly useful if we think not of an executive center an executive system , in which different areasof the frontal lobes are playing dissociablebut interconnectedroles. In our laboratories , we have carried out an analysis of a range of cognitive data in humans and primates , with the and have and . R. G. of assistance , J Taylor , tentatively idenBugmann Bapi, tified three principal subcomponentsof executive functioning for modeling . These are (1) establishing links between working memory representations purposes , , which could represent sensory stimuli, potential motor actions Farah & Rakic 1992 1987 Goldman or , rewards ; , , , Kimberg punishments ( schemas that level 1993); (2) creating, learning, and deciding among high , 1994; Pas (Grafman , but often flexible action sequences embody repeatable , 1993); and (3) incorporating affective evaluationsof sensoryevents singham or potential motor plans and using theseevaluationsto guide actions (Levine & Prueitt, 1989; Nauta, 1971). In this regard, many symptoms that frontal lobe patients have (some dorsolateral and some orbital, though it is often difficult to locate the lesion , 1996). Tranel, exactly) are seenas disturbancesof executive function (Lezak tests of different summarized and Benton 1994 Anderson, ( ) many aspectsof executive function: these included the WCST, used to test flexibility in ; verbal fluency, used to test both shifting between cognitive categories instructions to follow and ; and tests whose modeling ability spontaneity such as the Stroop test and of this in other is discussed book, chapters Tower of Hanoi.

VERBAL FLUENCY MODELING


The model in Levine and Parks (1992) of frontal effects on verbal fluency is : Levine, 1987; Levine derived from a previous model of the WCST (Leven & , & : Prueitt, 1989; Parkset al., 1992). While different in ways to be discussed the verbal fluency test and WCST have enough in common to suggest that . For example models of the two tests should sharesome features , both tests , and involve sustained attention, shifting of cognitive set, perseveration formation. category In the caseof verbal fluency, there are at least three possIble effects by , which frontal lobe damage can reduce the word count (Crockett, Bilsker Hurwitz , et al., 1986). The model neural network of Levine and Parks(1992)

366

in Dementia Applications

is constructedin sucha way that all three effectscan be produced in the network by weakeningof signalsfrom the samenode. The first effect is perseveration , which is a common result of frontal lobe , 1997, and Nauta, 1971). In the WCST, for damage(for reviews, see Fuster , the experimenterchangesthe reinforcementcriterion in the middle example of the task , from rewarding classifications basedon the color of the cards to classifications based on the . rewarding shape of the design on the cards Normal subjectsrespond to the change in reinforcement by changing their choices , whereasfrontally damagedsubjects tend to perseveratein making color choices despite receiving error messages . Since the subjects are not reinforcement the verbal receiving during fluency test, perseveration here takes on different forms. For example , frontally damagedsubjectsare more . Also, they are more likely to repeat words they have previously stated likely to say words that begin with letters that experimentersasked them , when they are successivelyasked to say words previously. For example with F C and L for one minute each , , , they might say C words beginning when they are supposedto say F words. The secondeffect is difficulty in following rules. Nauta (1971), for example , discussedseveralcasesof frontal patients violating rules of gamesthey were asked to play. In one game , subjectsare asked to move a stylus horizontally or vertically acrossa board to go through a maze . Frontally damaged often to take shortcuts the try patients by moving stylus diagonally, which is against the rules. In the verbal fluency task, frontally damaged patients show a higher incidence than do normals of using proper names , which they have been told not to do. The third effect is difficulty in paying attention to the task itself. Distractibility is a frequent effect of frontal lobe damage (Wilkins , Shallice : ,& . 1987 This relates to the functions of the prefrontal cortex ) McCarthy, in linking events across time (Fuster , 1997) and in forming goal-directed of behavior (Bapi & : Levine, 1994). In the verbal fluency task, the sequences minds of frontally damagedsubjectscould, for example , wander away from words with the letter C toward other words or entirely producing beginning irrelevant ideas . This task is particularly sensitive to distraction becauseit involves classification of words by a rule (opening letter) rather than by , dishes , parts of the body ). meaningful categoriesof experience(e.g ., animals 15.4 shows the neural network used to model frontal lesion effects Figure on verbal fluency (Levine & : Parks , 1992). The category nodes code the letters (C, F, and L) with which words begin. The Wisconsin Card Sorting network of Levine and Prueitt (1989) also included a feature layer, so that the feature and category nodes constituted a modified adaptive resonance module basedon the ideasof Carpenterand Grossberg(1987). In the current network, however, the feature layer is omitted becausethe only feature of . ( beginning letter) also definesthe categories consequence The node labeled "R ," for reinforcement has , multiple effects in this network . As in the WCST model of figure 15.2, frontal damagecorrespondsto

367

Parks& : Levine: Wisconsin Card Sorting and Verbal Fluency

Figure 15.4 Neural network model of verbal Ruencytest perfonnancefor frontal lobe com . control with nonnal subjects parisons

weakenedgain of signals from the reinforcement nodes to sensory (in this . This is suggestedby the fact that the frontal case , category and word ) nodes lobes are the main connecting link between sensory areasof the cortex and subcortical motivational areas of the limbic system and hypothalamus , 1997; Nauta, 1971). (Fuster First, the reinforcementnodesprovide tonic arousalin the form of random . The random signals are multiplied by a gain signals to all word nodes IX . If IX is too small called , background noise (external or internal , parameter to the network) tends to keep word node activation below the threshold , thus mimicking distractibility due necessaryfor words to be actually said . However, some words are typically spoken even to frontal lobe damage . To mimic that effect, we divided our word database with frontal damage " " words most of which had fewer than six letters, and "hard " into easy , words. The samegain parameterIXalso determinesthe strength of the other three sets of signals from the reinforcementnode. Two of these sets of signals go . One set of signals is to word nodes , but are more specific in their effects . The other consistsof inhibitory of those nodes that representproper names modulatory signals that influence the strength of self inhibition at the word nodes . Such self-inhibition is activated once a word node has exceededthe threshold for speech , to prevent the same word node from getting reactivated by signals from the node for the currently active category. This mimics the operations of a widely used neural network called the mechanism : Prueitt, 1989), : Grossberg , 1987; Levine & gated dipole field (Carpenter &

368

in Dementia Applications

INHIBITION

ON

SELF

GAIN

~ CATEGORIES 0 eeo QOOO , lJ . : ~ . . . . . , . ( ~ , \

+ TO ACTIVE CA ~

based on the principle of opponent processing , in which neurotransmitter creates a to disable . tendency depletion previously active nodes The fourth and final set of signalsfrom the reinforcementnode goes to the node for the active category, as indicated by which of the three letters (in our simulations , C, F, and L) the experimenter has instructed the subject to concentrateon. The three category nodes compete via lateral inhibition . If the signal to the active node is too weak, there can be perseverative ; for example , if the subject is asked to think of words beginning responses with C for one minute and then with F for another minute, the subject could continue to say some C words when he or she is supposedto say F words. Sucha perseverativetendency is strengthenedby positive feedbackbetween , which are similar to those category nodes and the correspondinghabit nodes utilized in the WCST model of figure 15.2. The word count numbersmay be varied, depending on the severity of the , by changing the total number of words in the database cognitive impairments " " with each letter and the proportion of " easy and hard " beginning " " words. For example , the frontally damaged network could be modeled to reproducemild to severely impaired deficits in verbal fluency. Owing to the perseverativetendency causedby positive feedbackbetween category and habit nodes , there were other parametervaluesfor which frontal damagehad an opposite effect. For those values , instead of reducing the word count, " " decreasingIXto the frontally damaged range causedthe network to say the samewords over and over. Our earlier model (Levine & : Parks , 1992) contained overestimations of the total number of words significant produced per minute for normal persons . However, the neural network model was successful overall, because it accuratelyreproducedthe magnitude of differences between normals and frontal lobe- damaged patients. In other words, the ' frontal lobe patients verbal fluency performancewas severelyimpaired when with that of normal controls. In the next section , our new simulations compared ' between Alzheimer s diseasepatients and elderly controls was designed to mirror updatedverbal fluency norms (Bentonet al., 1994), which reflect adjustments for age. Moreover, the updated model contains new neuroanatomical and neurotransmittercomputernodesthat enableus to model declinein verbal . fluency performanceas the patient passes through stagesof the disease MODELING EFFECTS OF A LZ HEI ME R' S DISEASEWITH VERBAL FLUENCY
In this section we first explain selected neuromechanisms that underlie ' Alzheimer s disease (AD ). Next , the rationale for the development of computer modules that reflect the neurotransmitter interactions with neuroana tomical regions such as the frontal and hippocampal areas are hypothesized . The computer model is then presented for verbal fluency that reflects decreases in word generation as a reflection of declining neural systems over a period of years .

369

Parks& : Levine: Wisconsin Card Sorting and Verbal Fluency

Neurobiological

and Neuroanatomical

Aspeds of Alzheimer ' s Disease

There is general agreementthat the major neuropathologicalchangesin AD : Van Hoesen are neurofibrillary tangles ( NFT , ) and senileplaques(Solodkin & Van et al. . and associates Hoesen 1993 1997) Hyman , , (Hyman, Arriagada, ) found from their anatomicalstudiesthat the hippocampalareais the first area to develop NFT, and that NFT accumulatein the associationcortices as the . A longitudinal positron emission tomography clinical severity increases of ) study early neuropsychologicaland metabolic changesin AD suggested (PET that degenerationand metabolic dysfunction first occur in the medial temporal area (Grady, Haxby, Horwitz , et al., 1988). This is followed by degenerationin the neocortical associationregions (frontal, parietal, and lateral temporal regions). In the study by Grady et al., episodic memory tasks such as the delayed verbal memory aspectsof the Wechsler Memory Scale , 1975) and the Buschke(1973) test become impaired one or more (Russell . Correspondingly, the a decline in verbal fluency performance before years AD subjects with initial impairment on the Wechsler Memory Scale and Buschketest were hypothesizedto have initial metabolic dysfunction in the medial temporal area , even though the small size of the medial temporal area relative to the spatial resolution of the tomograph did not allow precise and brain imaging studies imply . In summary measurement , neuroanatomical at function that loss of hippocampal proceeds a more rapid rate than loss of frontal function in AD . It is well known that AD dementia virtually always involves a deficit of : Decker , cholinergic innervation from the midbrain nucleusbasalis(Brioni & al. et 1981 Whitehouse Price Clark & : Bower 1992 Hasselmo 1997; , , ; , , ). , Therefore , we also add to our network a node representingcholinergic function . We assumethat cholinergic innervation affects both the hippocampal and frontal nodes of our network, but that the effect of cholinergic innervation on hippocampal activation is faster. This is based on experimental results showing that the basal forebrain is the major source of cholinergic input to some areas of the central nervous system (CNS), such as to the frontal cortex from the nucleusbasalisof Meynert (nBM ) and to the hippo. It sharescholinergic input also with intrinsic campusfrom the medial septum neurons in the cortex and with the tegmental nuclei in the thalamus , 1991). There may be a differential loss of cholinergic input that (Becker could account for the differential effect on function. It is not known whether the acceleratedloss of function in the hippocampusis secondaryto loss of cholinergic innervation from the medial septum or due to a more rapid neural . Inferencesfrom loss in the hippocampusthan in the associationcortices PET researchsuggestthat different areasof the nBM degenerateat different " " of metabolic and cognitive deficits (Grady rates , resulting in the sequence et al., 1988, p. 592). More recent neuropathological studies have yielded evidencethat hippocampusand entorhinal cortex show the earliest structural changesand dysfunction in AD , and that the frontal, parietal, and temporal

370

in Dementia Applications

areasfollow as the diseaseprogress es (BancherI Braak , Fischer, DeLacoste& White, 1993).

et aI ., 1993 ;

The details of cholinergic degenerationat the neuronaland synaptic level are modeled well in chapter 13. We concentrate instead on modeling at the " " " " " " system level, using simplified frontal, hippocampal, and cholinergic modules to simulate the effects of different stagesof AD on verbal fluency. in word count is relatively mild in the early stagesbut becomes The decrease at . The network used in our model of AD precipitous the intermediatestages is an and verbal fluency extension of the network of figure 15.4 that includes thesemodulesrepresentingdifferent brain areas . To model the effect of AD on verbal fluency, we have split the function of " the reinforcement node in figure 15.4 into frontal" and a "hippocampal" , with different time scalesof activity . In our expandednetwork, components shown in figure 15.5, the nonspecificarousal function previously attributed to the reinforcementnode is now attributed to the hippocampalnode. This is becausethe hippocampus is involved in setting the stage for the current situation, that is, orienting the organism to new stimuli (T eyler & : Di Scenna , 1986). The functions of rule generation are attributed to the frontal node. In the computer model of the verbal fluency performance , rule generation includes self-inhibition of nodes for previously stated words, inhibition of nodes for proper names , and excitation of nodesfor the active letter (the one the clinican is currently asking for ). Following a suggestion of Pennington (1994), based on an alternative connectionist model of Sloman and Rumel " " hart (1992), we propose that a part of the frontal cortex performs a gating or modulatory function on connections from this hippocampal node to stored semantic representations , in this case of previously learned words , 1986). (basedon the hippocampalindexing theory of Teyler and Di Scenna " " " " Frontal gating can either enhanceor depressa hippocampal activation of " a specificcategory or word node. The frontal" node sendsmodulatory signals that are multiplied by the hippocampal signals . The frontal signal is a high number for the active category and a very low (much less than 1) number . Self-inhibitory signals at for word nodes representing proper names word nodes are used to prevent repetitions, as in the network of figure 15.4. " " The frontal cortex decides which word representationsto gate positively or negatively.
Neuropsychological

Implications

Memory tasks with the greatest hippocampal involvement should be affectedthe most in the early stagesof AD . In particular, episodic memory is affected earlier in the course of the diseasethan is semanticmemory (Zec,

371

Parks & : Levine : Wisconsin CardSortingandVerbalFluency

WORDS 000 cat face laugh LETTERS @ CF L Slow CATEGORIES @ animals foods actions HABITS A Ch SOURCE IPPOCAMPAL Fast

1993). Recall the distinction made by Tulving (1972): semantic memory involves recall of stored facts and relationships(e.g ., " Springfield is the capital of Illinois " ), whereasepisodic memory (for review, seeWheeler, Stuss : ,& " Tulving , 1997) involves recall of specificevents (e.g ., I ate pizza for dinner " yesterday ). While the verbal fluency test has one episodic component rememberingwhich letter the experimenteraskedfor it is predominantly a semanticmemory task. The full network of figure 15.5 was designed to allow modifications for the possibility of modeling at least one task that primarily involves episodic memory. The task is the Rey Auditory Verbal Learning test, which involves recall from a list of fifteen words presented five times (Powell, Cripe, & : Dodrill , 1991; Spreen& : Strauss . 1998 Our is , ) laboratory currently working on the Rey Auditory Verbal Leaning test model that usesa similar database as seenfor verbal fluency performancein table 15.2. Preliminary versions of our episodic task model include intrusion errors (words totally unrelated to the list recall), in addition to perseverationerrors.

372

in Dementia Applications

Table 15.2 Partial representativeword databaseshowing initial letter and semanticcategory


SemanticCate~ory Foods Body parts or clothes Animals House objects Actions Adjectives Proper names C words F words L words

cheese cherry. . . chincoat. . . catcow. . . chaircup... callclap... coldcreepy ... CarlCindy. . .

fishAg... foot face. . . frog fox ... Roorfan.. . fall Ale... famous fatal. . . France Fred. . .

lettuce lobster... leg lip . . . lamblion . . . lamplight . . . laughleap. . . long lame... LisaLarry...

Representativesimulation of verbal fluency results are shown in figure " " 15.6. The deterioration parameter measuresthe amount of loss of cholinergic function, so that the abscissadescribesprogressive stages of AD . Note that verbal fluency declines more gradually for mild AD , then drops " " . Verbal fluency is precipitously at an intermediate stage of the disease impaired in some studies comparing elderly controls to patients with mild AD (Barr & : Brandt, 1996; Hart, Smith, & : Swash , 1988; Miller , 1984; Ober, Dronkers, Koss , et al., 1986). Moreover, verbal fluency is sensitive to impairments seen in other clinical groups such as Huntington ' s disease , et al., 1994). Other researchhas suggested that (Monsch, Bondi, Paulsen discriminates fluency categorical patient groups better than the letter fluency : Chan test usedin our experiments(for review, seeSalmon& , 1994). ' the of , both severity patients neuropathologic changes Depending on frontal lobe and AD patients have deficits in executive functions when compared with control groups. More specifically , these patients have difficulty with concurrent manipulation of information such as set shifting and selfmonitoring , Sala , Papagno , et al., 1997; Bondi, Monsch, Galasko , (Baddeley : Albert , 1995; Parks : Wilson, 1993; Salmon et al., 1994; Lafleche& , Zec, & , 1997; Zec, Parks , Gambach , et al., 1992). Caution should be used as to rigid use of the neural network model presentedhere. We modeled some of the more obvious behaviors , neuropathologic findings, and neurochemicalsystems . Important aspects of human functioning were not fully addressed , namely the role of motor networks and visuospatial integration. On the , a careful analysis of various human biological systems has positive side accurately reproduced selected cognitive functions with the assistanceof innovations in computer programming.

ACKNOWLEDGMENTS
The information in this chapter contains significant adaptations and reproductions " from Parallel Distributed Processing and Neuropsychology: A " Neural Network Model of Wisconsin Card Sorting and Verbal Fluency by Review Parks et al., 1992, Neuropsychology , Plenum Publishing Corporation, . A representationof a psychological test used in this used with permission

373

: Wisconsin Card Sorting and Verbal Fluency Parks & : Levine

, ' 16 . . . , . . , , 12 . FLUENCY : P ER ( MINUTE )8 ~ , . 4 ', : .. 24 :6 : , . ,8


NORMAL TO MILD

SEVERE MODERATE

ACh DECLINE IN YEARS


Figure 15.6 Decline in verbal Ruency test performance model over a period of years in . This graph shows the various stagesof the disease Alzheimer ' s disease patients progressionand A . Ch ( ) cholinergic dysfunction

, Psychological chapterwas reproducedby specialpermissionof the publisher AssessmentResources , Inc., 16204 North Florida Avenue, Lutz, FL 33549, Horn the Wisconsin Card Sorting test and the WisconsinCard Sorting Test Manual by Dr. Robert Heaton, et al., copyright 1981. Further reproduction is prohibited without permission of Psychological AssessmentResources , Inc. Code for the Wisconsin Card Sorting and verbal fluency models is being madeavailableat http :// www .uta.edu/ psychology/faculty/levine.

REFEREN CFS
. (1977 Amari , 5.-1 ) . Dynamics of pattern fonnation in lateral-inhibition type neural fields. Bio-

- 87. , 27 , 77 Cybernetics logical

-taskperfonnance : Spinnler in dysexecutive , A , Sala . S. D., Papagno , C., 6 , H. (1997 ). Dual Baddeley - 194 with a frontallesion . Neuropsychology andnondysexecutive . , 11 , 187 patients : Jellinger . A (1993 Bancher . H., Fischer , P., 6 ,K , C., Sraak ). Neuropathological stagingof Alz' s disease helmer lesions andinte Uectua I status in Alzheimer ' s andParkinson . Neuroscima patients Letters . - 182 , 162 , 179 . (1994 . S., 6 : Levine ,R , D. S ). Modelingthe role of the frontallobesin performing Bapi sequential - 1180 . Neural tasks . I. Basic strudureandprimacy effects Networks . ,7 , 1167 . S., 6 : Levine ,R , D. S. (1997 ). Modelingthe role of the frontallobesin performing Bapi sequential . Neural tasks . o. Classification of sequences Network World , 1(97 ), 3- 28.

374

Applications in Dementia

-list generation Barr : Brandt deficitsin dementia . ] ouma -l of Clinical and , A, & , J. (1996 ). Word - 822 . , 18 , 810 Erperimental Neuropsychology Becker . E. (1991 ' s Disease : The cholinergic ,R ). Therapyof the cognitivedeficitin Alzheimer . Becker . In R & :E . Giacobini .), Cholinergic basis (Eds (pp. 1- 24). Boston system for Alzheimer therapy : Birkhiuser .
. Neuropsychologill Benton . A . t . (1968). Di Herentialbehavioral effeds in frontal lobe disease , 6, 53- 60. ., & : Sivan Benton . A t ., Hamsher ,K , A . B. (1994). Mul Hling , 3rd ed. Iowa U RlllphRSilluaminaHon . City : AlA Associates Bliss : Lomo, T. (1973). Long-lasting potentiation of synaptic transmission in the , T. V. P., & dentate area of the anaesthetized rabbit following stimulation of the perforant path. Journalof

- 356 . , 331 (London ), 232 Physiology

: Delis Bondi , M. W., Monsm , A V., Galasko , D., Butters , N., SalmonD . P., & , D. C. (1994 ). Preclinical - 384 markers of dementia of theAlzheimer . Neuropsychology . , 8, 374 cognitive Type ' Brioni . D., & : Decker M. W. (Eds .). (1997 ~ l treatment s disease . New ,J ). Phimnacologi of Alzheimer -Liss . York : Wiley Buschke for analysis of memoryand learning . Journal , H. (1973 ). Selective of Verbal reminding and Verbal Behavior 12 543 550 . , , Learning : Grossberg architecture for a self - organizing , G. A., & ) . A massively , S. (1987 Carpenter parallel neuralpatternrecognition machine . Computer Vision , Graphics , and Image , 37 , 54Processing 115 . Carter : Cohen and facilitationeffectsduring , C. 5., Mintun , M., & , J. D. (1995 ). Interference - 272 selective attention : An H2 150 PETstudyof Strooptaskperformance . Neuroimage . , 2, 264 : An allostericmembrane . Haroty , J.-P. (1981 ). The acetylcholine Changeux receptor protein Lectures 75 . , 85- 254 Cohen : McClelland :A es , J. D., Dunbar , K., & , J. L. (1990 ). On the controlof automatic process - 361 distributed account of theStroopeffect . Psychological Review . , 97 , 332 parallel processing -Schreiber : Servan : A connectionist Cohen , J. D., & , D. (1992 , cortexand dopamine ). Context to behavior andbiologyin schizophrenia . Psychological Review , 99 , 45- 77. approach : Eichenbaum Cohen . Cambridge , N. J., & , H. (1993 , amnesia , andthehippocampal ). Memory , system MA: MIT Press . Crockett : Kozak . (1986 of , D. J., Bilsker , D., Hurwitz ' T., & . J ). Clinicalutility of threemeasures in neuropsychiatric frontal lobe dysfunction . lntmrational Neuroscience 30 , , Journal of samples - 248 241 . Darnasio comments . In H. S . Levin : A L. Benton , A R. (1991 , H. M. Eisenberg ). Concluding ,& - 407 .), Frontal lobe anddysfunch "on(pp. 401 . (Eds ). NewYork: Oxford UniversityPress function ' error A : Darnasio R . 1994 . Dtsc Rrles Emotion reason and the human brain . New York : Grosset , , , ( ) Putnam . -P. (1989 Dehaene : Changeux modelof prefrontal cortexfunctionindelayed , 5., & .J ). A simple - 261 tasks . Journal Hve Neuroscience . ,1 , 244 of Cogni response Dehaene : Changeux cardsortingtest : Theoretical and , 5., & . J-P. (1991 ). TheWisconsin analysis in a neural . Cerebral . network Cortex , 62- 79 ,1 modeling DeLacoste : White in Alzheimer 's disease , M., & , C. L. (1993 ). The role of corticalconnectivity : A reviewandmodelsystem . Neurobiology . , 14 , 1- 16 ofAging pathogenesis

375

Parks : Wisconsin & : Levine CardSortingandVerbalFluency

Dias, R . , Robbins : Roberts , T. W., & , A . C. (1996). Dissociation in prefrontal cortex of affective and attentional shifts. Nature , 380 , 69- 72. Fuster , J. M . (1997 ) . Theprefrontalcoriu : Anatomy , physiology , and neuropsychology of thefrontal lobe . 3rd ed. New York: Raven . -Rakic Goldman , P. S. (1987 ) . Circuitry of primate prefrontal cortex and regulation of behavior .), Handbookof physiology by representationalmemory. In F. Plum (Ed , Vol . 5 (pp. 373- 417) . Bethesda , MD : American Physiological Society. -RakiC Goldrnan , P. s. (1992). Working memoryandthe mind. Scientific American , September , 73- 79. Grady, C. L., Haxby, J. V., Horwitz , B., Sunadaram , M ., Berg . P., , M . B., Friedland ,R , G., Schapiro & : Rapoport, S. I. (1988). Longitudinal study of early neuropsychologicaland cerebralmetabolic changesin dementiaof the Alzheimer type. Journalof Clinical and ErperimmtalNeuropsychology , 10 , 576- 596. Grafrnan , J. (1994). Alternative frameworks for the conceptualizationof prefrontal lobe functions . In F. Boller & : J. Grafrnan ( Eds .), Handbookof neuropsychology , Vol . 9, (pp. 187- 201). Amsterdam : Elsevier . ). A neural theory of punishment and avoidance . I. Qualitative theory. Grossberg , S. (1972a Mathtmlltical Biosciences 15 , 39- 67. . II. Quantitative theory. ). A neural theory of punishment and avoidance Grossberg , S. (1972b Mathtmlltical Bioscimces 15 , 253- 285. , short term memory, and constanaesinreverberating Grossberg , S. (1973). Contour enhancement neural networks. Studies in AppliedMathematics52, 213- 257. in mind and brain. Dordrecht, Netherlands : Reidel . Grossberg , S. (1982). Studies and natural intelligence . Cambridge . , MA : MIT Press Grossberg , S. (1988). Neuralnetworks : Mingona, E. (1985). Neural dynamicsof perceptualgrouping: Textures Grossberg , S., & , boundaries . Perception , and emergentsegmentations and Psychophysics , 38, 141- 171. Hart, S., Smith : Swash , C. M ., & , M . (1988). Word fluency in patients with early dementia of Alzheimer type. BritishJournalof Clinical Psychology , 27, 115- 124. Hartline, H. K., & : Ratliff, F. (1957 ) . Inhibitory interactionsof receptor units in the eye of Limulus . , 40, 351- 376. Journalof Gtntral Physiology Hasselmo : Bower, J. M . (1992). Cholinergic suppression , M . E., & speafic to intrinsic not afferent fiber synapses in rat pirlform (olfactory) cortex. Journalof Neurophysiology , 67, 1222- 1229. Heaton , R. K., Chelune , G. J., Talley, J. L., Kay, G. G., & : Curtiss Card Sorting , G. (1993). Wisconsin testmanual : Revised and expanded . Odessa . , FL: PsychologicalAssessment Resources Hebb, D. O. (1949). Theorganization . New York: Wiley . of behavior Hestenes , D. O. (1992). A neural network theory of manic - depressive illness . In D. S. Levine & : S. J. Leven (Eds .), Motivation, emotion , and gMI direction in neural nehoorks(pp. 209- 257 ). Hinsdale . , NJ: Erlbaum Hestenes in mental disorders , D. O. (1998). Modulatory mechanisms . In D. J. Stein & : J. Ludik .), Neuralnehoorks in psychopathology (Eds (pp. 132- 164). Cambridge , U.K.: Cambridge University Press .

Hull, C. L. (1943 . New York : Appleton . ). Principles of behavior : Darnasio , P. V., Van Hoesen . (1993 , G. W., & Hyman , B. T., Arriagada , A. R ). Memoryimpairment in AlzheimersDisease : An anatomical . In W. R. Parks R . F. Zec : R . S. , ,& perspective 's Disease Wilson (Eds .), Neuropsychology andotherdtmtntill S(pp. 138 - 150 of Alzheimer ). New York : Oxford University Press .

376

Applications in Dementia

: Squire . 5., Shimamura , M., & , L. R. (1989 ,J , A P., Kritchevsky ). CognitiveImpalment Janowsky . Behavioral andits relevance to hwnanamnesia Neuroscience , 103 , followingfrontallobedamage 548 - 560 . R Hon : The biochmlicalcontrol of neuron Rl : Levitan Kaczmarek . L. K , & , I. B. (1987 ). Neuromodul : Press . . New York Oxford University udtability facilitation in neuron es of the abdominal : Tauc Kandel , E. R., & , L. (1965 ). Heterosynaptic ganglion . Journal of Aplysi Rdepil Rns , 1- 27. (London ), 181 of Physiology ' of mediation : Haywood , Y. v ., & , H. C. (1998 ). Two waysto elaborate Vygotskys concept Karpov . American , 53 , 27 36. Psychologist : Farah of cognitiveimpairments , M. J. (1993 ). A unifiedaccount following Kimberg , D. Y., &
: The role of working memory in complex frontal lobe damage , organized behavior. ] ournRlof - 428. : 122 411 General , , Psychology Erperimental

: W. a Kolb, B., & : Whishaw, I. Q . (1995). Fundamentals , 4th ed. New York / humanneuropsychology . H. Freeman
' s disease . Neuropsychology Lafleche : Albert , M . S. (1995). Executivefundion deficits in mild Alzheimer , G., & , 9, 313- 320. : Levine, D. S. (1987). Effectsof reinforcementon knowledge retrieval and evaluation Leven , S. J., & on neural networks . In IEEEfirst inttrn Rtionalconference , Vol . 2 (pp. 269- 279). San Diego: IEEE /ICNN .

. to neural andcognitive . HiUsdale : Erlbaurn Levine , NJ , D. 5. (1991 ). lntrodudion modeling . In J. Reggia of the prefrontalexecutive Levine , , D. 5. (1996 ). Modelingdysfunction system - 439 : . (pp. 413 .), Neural E . Ruppin : R. Berndt(Eds , NJ ,& ). RiverEdge of braindisorders modeling . World Scientific . in a networkmodel lesioneffects on verbalfluency Levine : Parks , R. W. (1992 , D. 5., & ). Frontal : Neh Dorks In lntmrational onNeural , NJ , Baltimore , Vol. 2 (pp. 39- 44). Piscataway JointConference IEEE . in neural . 5. (1993 andtheoretical issues : Prueitt Levine , D. 5., Parks , R. W., & ,R ). Methodological . lntmrational networkmodels of frontalcognitivefunctions , 72 , 209 Jounurl of Neuroscience 233 .
: Novelty and Levine : Prueitt, P. S. (1989). Modeling someeffectsof frontal lobe damage , D. 5., & - 116. Networks 2 103 . Neural , , perseveration . assessment , 3rd ed. New York: Oxford University Press Lezak , M . (1996). Neuropsychological Hons in man Luria, A . R. (1980). Highercorlicalfunc , 2nd ed. New York: BasicBooks. : Knight, R. T. (1996). Impaired retrieval from remote memory , J. A ., Gershberg , F. B., & Mangeis . Neuropsychology in patients with frontal lobe damage , 32- 41. , 10 : Schacter , D. L. (1997 ) . The neuropsychology of insight: Impaired awareness McGlynn, S. M ., & Annals of defidts in a psychiatric context. Psychiatric , 27, 806- 811. Miller , E. (1984). Verbal fluency as a function of a measureof verbal intelligence and in relation to different types of cerebralpathology. BritishJournalof Clinical Psychology , 23, 53- 57. 9, 90- 100. Milner, B. (1963). Effectsof different brain lesionsin card sorting. Archivesof Neurology .), : K. Akert (Eds Milner , B. (1964). Someeffects of frontal lobectomy in man. In J. M . Warren & Thefrontal granularcorlexand behavior (pp. 313- 334). New York: McGraw-Hill . Milner , B., & : Petrides , M . (1984). Behavioral effects of frontal-lobe lesions in man. Trendsin Neurosciences , 7, 403 407.

377

Parks : Wisconsin CardSortingandVerbalFluency & : Levine

Mishkin : Ba chevalier , M., Malamut andhabits : Two neural , B., & , J. (1984 . In ). Memories systems G. Lynch : N. Weinberger .), Neurobiology , J. McGaugh tmdmemory ,& (Eds (pp. 65- 77 ). of / taming NewYork : GuilfordPress . Monchi : Taylor , 0 ., & , J. G. (1995 the basal ). A modelof the prefrontal loop that includes ganglia in solvingthe recency task . In Worldcongress on neural networks , Washington , DC, Vol. 3 : Erlbaum . , NJ (pp. 48- 51 ). Hillsdale Monsch , A U., Bondi , M. W., Paulsen , J. S., Brugger : Swenson , P., Butters , N., SalmonD. P., & , 's M. (1994 of category andletterAuency in Alzheimer 's disease andHuntington ). A comparison disease . Neuropsychology , 8, 25- 30. -task interference and dual Moscovitch , M. (1994 effectsat retrievalin ). Cognitiveresources normalpeople : The role of the frontal lobesand medialtemporal cortex . Neuropsychology , 8, - 534 524 . of the hontallobe : A reinterpretation Nauta , W. J. H. (1971 . Journal ). Theproblem of Psychiatric - 187 Rt Se Rrch . , 8, 167 Ober : Friedland . P. (1986 , B., Dronkers , N. D., Koss , E., Delis Hornsemantic , D. C., & ,R ). Retrieval -type dementia . Journal R I andE memoryin Alzheimer , 8, of CliniC :rptrimental Neuropsychology - 92. 75 Parks . W., & : Cardoso distributed ,R and executive , J. (1997 : ). Parallel processing functioning Towerof Hanoineural networkmodelin healthycontrols andleft hontallobepatients . International - 240 . , 89 , 217 Journal of Neuroscience Parks . W., Crockett : McGeer ,R modelof corticalorganization , D. J., & : , P. L. (1989 ). Systems Positronemission and neuropsychological test performance . Archives R I tomography of CliniC - 349 . , 4, 335 Neuropsychology Parks . W., Crockett ,R . L , Zec , D. J., Tuokko , H., Beattie , B. L., Ashford , J. W., Coburn ,K , R. F., " " Becker : McGeer , R. E., McGeer , P. L , & , E. G. (1989 ). Neuropsychological systems efficiency andpositronemission . and I CliniC R 1 Neurosciences 269 , Journal , of Neuropsychiatry tomography . 282 Parks . W., Levine ,R , D. S., Long , D. J., Dalton , I. E., Weingartner , H., Fedio , P., , D. L , Crockett .R Matthews . , Coburn : Becker ,J , K. L , Siler distributed , G., & , R. E. (1992 ). Parallel processing andneuropsychology : A neural networkmodelof Wisconsin CardSortingandverbal8uency . Review 3 213 233 . , , Neuropsychology Parks ., Crockett , R. W., Long , D. S , D. J., Dalton , I. E., Zec , R. F., Coburn , K. L , , D. L , Levine Siler , G., Nelson : McGeer , M. E., Bower , J. M., Becker , R. E., McGeer , E. G., & , P. L (1991 ). Parallel distributed andneural networks : Origins andcognitivefunctions processing ., methodology - 214 . Internationa I Journal . , 60 , 195 of Neuroscience 's disease Parks . W., Zec . F., & : Wilson .). (1993 ,R ,R and , R. S. (Eds ). Neuropsychology of Alzheimer other demen Has . NewYork : Oxford University Press . . (1993 andvoluntary action . Oxford , R. E .: Oxford University ). The , U.K frontallobes Passingham Press . Pavlov , I. P. (1927 . ) . Conditioned refle :usV. Anrep(transLondon: Oxford UniversityPress functionof the prmontal cortices , B. F. (1994 : Implications ). The workingmemory Pennington for developmental andindividual differences in cognition . In M. Haith : , J. Benson , R. Roberts ,& - 289 B. F. Pennington .), Future oriented esin development : University (Eds (pp. 243 ). Chicago process of Chicago . Press Perret , E. (1974 of habitual in verbal ). Theleft hontallobeof manandthe suppression responses - 330 behavior . Neuropsychologia . , 12 , 323 categorical

378

Applications in Dementia

: Evidence from investigations of the Petrides , M. (1994 ). Frontallobesand workingmemory In F. . nan . Hlmdbook in nonhuman . Boller & : Gra& effects of corticalexcisions J (Eds ), of primates . : Elsevier , Vol. 9, (pp. 59- 82). Alnsterdam neuropsychology of brainimpairment with the Rey : Dodrill , C. B. (1991 Powell J. B., Cripe , L I., & ). Assessment measures . Archives : A comparison with otherneuropsychological Test AuditoryVerbalLearning 6 241 . 249 Clinical , , of Neuropsychology . In K. H. Pribram & : : Executive of the brain Pribram , K. H. (1973 ). The primatefrontalcortex - 314 : Academic .), Psychophysiology A R. Luria (Eds (pp. 293 ). New York of thefrontal lobes . Press . Hillsdale : Holonomy andsfn4cture . H. (1991 , Pribram , K ). Brainandperception infigurAlprocessing : Erlbaum . NJ : Francis . .recutive . Bristol PA: Taylor& Rabbitt , P. (1998 ). Methodology function offrontalande deAcits andcriticalreviewof neuropsychological : Wolfson Reitan , R. M. & , D. (1994 ). A selected - 198 . Review . Neuropsychology andthe frontallobes , 4, 161 . , Lehman . A W., & : Stilson , D., W. (1980 , R ). The utility of the Robinson , A L , Heaton , R. K . Journal frontallobelesions andlocalizing Wisconsin CardSortingTestin detecting of Consulting - 614 . andClinical , 48 , 605 Psychology . New York : Oxford University nthDorks andbrain ., & RoDs : Treves ,E , A. (1998 ). Neural function . Press of complex method for the assessment Russell . W. (1975 ,E ). A multiplescoring memoryfunctions . - 809 andClinical . Journal , 43 , 800 Psychology of Consulting . In J. D. Brioni : of Alzheimer ' s disease features ,& SalmonD. P. (1997 ). Neuropsychological 's diStQ : Moltcul Rr andneurobiologiall St treafmtnt M. W. Decker .), Pharmacological (Eds of Alzheimer - 147 : Wiley-Uss . Rtions ). New York (pp. 129 found with Alzheimer ' s disease deficits associated . P., & : Chang SalmonD ). Semantic memory , A. S. (1994 : in honor and .), Neuropsychological . In L S. Cennak (Ed Essays of memory cognition uplorations . : Plenum Press BuHm(pp. 61- 76 ). New York of Nelson memories in distributed interference : Rumelhart , D. E. (1992 Sloman , S. A , & ). Reducing through to . M. Shiffrin(Eds .), Fromlt Rming : R . In A F. Healy , S. M. Kosslyn theory , & gating episodic . . Hillsdale : Erlbaum K tes . 227 248 in honor Willi Rm & : connectionist , ) NJ (pp of theoryEssays of AlzW. and functional . Van G. 1997 A & : Solodkin Hoesen , ) Neuropathology , , ( anatomy 's treatment .), Plumnacological : M. W. Decker 's disease . In J. D. Brioni heimer ,& (Eds of Alzhtimer : Wiley. di St Q ). NewYork St ( 151 177 pp. -process : I. Temporal . L, & : Corbit , J. D. (1974 ). An opponent Solomon ,R theoryof motivation 145 . Review . Psychologiall of affect , 119 , 81 dynamics . New York tests 2nd ed : . (1998 : Strauss ,E , 0 ., & ). A compendium of neuropsychological Spreen . OxfordUnivenity Press to anterior Uice : Alexander Stuss , M. P. (1995 ). A multidisciplinary , D. T., Sha , T., & approach - 211 . functions . Annals attentional , 769 , 191 of Sciences of theNewYorkAcademy . Bthlwioral ., & : Di Scenna , P. (1986 ). The hippocampal , T. J memoryindexingtheory Teyler - 154 . Neuroscience , 100 , 147 of the concept of "executive : Benton Tranel , A (1994 ). Development , S. W., & , D., Anderson " . . In F. Bo Der & : . Grafman frontal lobes to the function andits relationship (Eds ), Hlmdbook J of : Elsevier . 148 . Amsterdam . 125 Vol. 9 , , ( ) neuropsychology pp

379

: Levine: Wisconsin Card Sorting and Verbal Fluency Parks&

. , Moscovitch ,A K : Winocur , M., & Troyer , G. (1997 ) . dusteringandswitdtingastwo components of verbalfluency : Evidence from youngerandolder healthyadults . Neuropsychology , 11 , 138 - 146 . andsemantic ). Episodic . In E. Tulving& : M. Donald Tulving , E. (1912 son(Eds .), Organization memory - 403 : Academic (pp. 381 Press . of memory ). NewYork Von derMalsburg - organization of orientation ). Self sensitive cellsin thestriate , C. (1913 cortex . . , 85- 100 Kybtmttik14 Welsh : Pennington , M. C., & , B. F. (1988 frontal lobe functionin children ). Assessing : Views fromdevelopmental . Developmental - 230 . , 4, 199 psychology Neuropsychology Wheeler : Tulving , M. A , Stuss , D. T., & ). Towarda theoryof episodic : The , E. (1991 memory - 354 frontallobes andautonoetic consciousness . PsychologiC R I Bulletin . , 121 , 331 Whitehouse , P. J., Price , D. L , Clark , A W., Coyle : Delong , J. T., & . (1981 disease ). Alzheimer , M. R : Evidence for selective lossof dtolinergicneurons in the nucleus basalis . Annals of Neurology - 126 . , 10 , 122 Wilkins : McCarthy , A. J., Shallice , T., & . (1987 ,R ). Frontallesionsand sustained attention . - 365 . , 25 , 359 Neuropsychologia . F. (1993 Zec ,R ). Neuropsychological in Alzheimer 's disease . In R . W. Parks , R. F. functioning 's Di :R . S. Wilson(Eds .), Neuropsychology Zec ,& St Rse andother Dementias of Alzheimer (pp. 3- 80). NewYork : Oxford University Press . . F., Parks . W., Gambach Zec ,R : Vicari ,R . (1992 , J., & ,S boardsystem : An innovative ). Theexecutive to cognitive -behavioral rehabilitation in patients with traumatic braininjury. In approach C. J. Long& - 230 :LK . Ross .), Handbook trauma : Acute (Eds CRrt to recovery of head (pp. 219 ). New York : Plenum Press .

380

Applications in Dementia

16

in Abnormalities Semantic Network ' s Disease Patients with Alzheimer Agnes S. Chan, David P. Salmon, and Nelson Butters

The semanticnetwork model assumes that our general fund of knowledge is organized as a complex network consisting of concepts or representations that are related through serial or parallel associationsor both (Collins & : . the Rumelhart & : McClelland 1986 Within network concepts 1969 , ; , , ) Quillian that sharemany attributes are more highly associated , or more closely : Loftus, 1975). For related , than those that share few attributes (Collins & the is associated with the , dog strongly concept cat because example concept -legged animals four share such attributes as creatures , , they being living in In and small size . contrast the domestic pets, meat eaters , , furry , relatively , at most, since they do not conceptsdog and sky are only weakly associated shareany obvious salient attributes. Psychologicalevidence for the associative nature of semanticknowledge has been provided by numerous studies : Quillian , 1969; using such diverse tasks as sentenceverification (Collins & : Schva : Loftus, 1975; Anderson, 1983), semanticpriming (Meyer & Collins & : Smith, 1973), and analogies neveldt, 1975), typicality ranking (Rips, Shoben ,& (Rips et al., 1973). Although the semanticnetwork model provides a functional view of how knowledge may be organized in the human mind, it does not provide a clear model of how semantic memory might be representedin the brain. Some , however, in developing neural progresshas beenmade over the past decade network models of semanticmemory by studying the effects of brain damage ice, 1988). Studies on the integrity of semantic knowledge (e.g ., Shall of patients with progressive anomia and impaired word comprehension , for example , suggest that semantic knowledge is stored in a distributed fashion in the association cortex of the temporal, frontal, and parietal : Warrington , 1988, 1990; Warrington & : Shallice lobes ( McCarthy & , 1984). This hypothesis has been supported by studies of other neurologic patient : Hodges & : Patterson , 1995) and by populations (e.g., semantic dementia in nonhuman see Fuster studies 1995 , , for a comprehensive ( primates analogous review). A neural network model that postulates that semantic knowledge is stored in a distributed fashion in the cortical associationareashas important

implications for the expected nature of the cognitive impairment associated with Alzheimer' s disease(AD ). AD is a progressivedementing disorder that is characterized by neuron loss, synapseloss, and the development of characteristic , neuneuropathological abnormalities (e.g ., neurofibrillary tangles ritic plaques , amyloid deposition), primarily in the hippocampal formation and the cortical association areas of the frontal, temporal, and parietal lobes (Terry & Katzman , 1983; Terry, Masliah, Salmon , et al., 1991). Because the diseaseprominently affects those brain regions that are thought to be integrally involved in the neural network underlying semanticknowledge, a disorganization of this knowledge might be expected as the disease es. progress Consistent with this notion, one of the primary clinical manifestations early in the course of AD is a languageimpairment that some investigators have attributed to a loss of semantic knowledge and a breakdown in the organization and structure of semanticmemory. Evidencefor this disruption ' includes AD patients particularly poor performanceon verbal fluency tasks in which they must generate exemplars from a specific semantic category , Granholm, Salmon (Butters , et al., 1987), a propensity to produce semantic errors on tests of confrontation naming (Bayies & T omoeda , 1983; Hodges, Salmon , & Butters , 1991), and consistency in producing errors for a particular semantic concept regardless of the mode of access and output (e.g., naming, sorting by category, word -picture matching: Chertkow & Bub, 1990; Hodges , Salmon , & Butters , 1992). These deficits in semantic become more over time (e.g ., Hodges memory , Salmon , & pronounced Butters 1990 , ), presumablyas the pathological abnormalitiesof AD progress and the neocortical associationareasbecomemore involved (Braak & Braak , 1991). In this chapter we describea number of recent studies that have directly explored the impact of AD on the integrity of the semanticnetwork. These studies employed various multidimensional , graphic scaling techniques , Shepard (Romney , & Nerlove, 1972; Shepard , Romney , & Nerlove, 1972; & Hutchinson 1986 to model the , ) Tversky organization of semanticmemory in AD patients to see how this organization might be altered by the neocortical deterioration that occursin the disease . The underlying hypothesis for these studieswas that deterioration of the neocortical sites of storage of semanticrepresentationsand associationswould lead to a gradual alteration of the associative network. This hypothesized relationship between cortical degradation and semanticnetwork deterioration in AD is reminiscent " of the observed effects of artificial " lesions imposed on computational " neural network models . As Rumelhart and McClelland (1986) stated , performance is characterizedby a kind of graceful degradation in which the 's system performancegradually deteriorates as more and more neural units are destroyed .. . This kind of graceful degradation is characteristicof such ' " global degenerativesyndromesas Alzheimer s disease. . . (p. 134).

382

Applications in Dementia

'S I ME R THESf MAN TIC NETWORKINPATIENTSWITH ALZHE DISEASE


Over the past five years , a number of studies &om our laboratory have used multidimensional graphic scaling analysesto explore the underlying structure . In of data generatedby AD patients on various semanticmemory tasks , these data analysis techniquesgenerate a graphic representationof general the data based on some measurementof the relative degree of association (i.e., closenessor proximity ) between concepts in semantic memory. The resulting models embed the conceptsunder study within a coordinate space where the distancesbetween points are assumedto reflect the psychological networkor cognitive pro.rimity between the respectiveitems in an internal semantic & Loftus 1975 Collins & 1969 Collins , ; Anderson, , ; ( Quillian map 1983). In an initial study (Chan , et al., 1993a , Butters, Paulsen ), we comparedthe " " organization of the semantic networks for the category animals in AD . In addition, we examined patients and elderly normal control (NC ) subjects ' the semantic networks of patients with Huntington s disease(HD ), a progressive dementing disorder that predominantly affects subcortical (i.e., the . The proximity between basal ganglia), rather than cortical, brain structures of their responses &om the for each was estimated subject sequence concepts on a category fluency task in which they generatednamesof animalsfor one that through spreadingactivation within thesemanminute. It was assumed tic network (Collin & Loftus, 1975; Meyer & Schvaneveldt , 1975), highly animalswould tend to be produced closer together within a subassociated ' . Therefore ject s sequentialresponsesthan those that are not highly associated two between that intervene number of items the , target words, corrected for the number of words produced , provided an estimate of the semantic proximity of the target words. The proximity scores were subjected to multidimensional scaling ( MDS ) and clustering analyses which of semantic the a model of knowledge for eachgroup organization generated . of subjects Although MDS analysesresulted in cognitive maps for all three groups that were best representedby two dimensions , domesticityand size (figure 16.1), the cognitive map of the AD patients differed consider ably &om those of individual in of the classification terms and HD of the NC subjects patients . For example , but not NC , in the cognitive map of AD patients exemplars , or HD subjects , bear appearedin the spacerepresenting domestic animals . In addition, the and zebraappearedin the spacerepresentingsmall animals cluster analysesrevealed two global clusters for the NC subjects and HD , whereasthree essentially patients which were interpretable as wild and domestic AD for the were derived clusters patients. uninterpretable The close similarity in the cognitive maps and clusters of animal names generatedfor HD patients and NC subjectsis of particular interest because the HD patients produced as few exemplarsin the fluency task as the AD

383

' s Disease : Butters: SemanticNetwork Abnonnalities in Alzheimer Chan ,& , Salmon

Small

Wild

DAT

HD

. This result is consistentwith previous studies that suggestthat the subjects poor fluency performanceof HD patients is due to a general retrieval deficit rather than to a breakdown in the organization of semanticmemory (Butters et al., 1987), and indicates that the abnormal cognitive maps of the AD patients are not likely to be an artifact of poor fluency performanceper se. The study describedabove was the first to demonstratewith MDS techniques that the organization of semanticknowledge becomesabnormal as a of AD , a cortical degenerativebrain disorder. However, methconsequence odologicallimitations of this study precludeda full exploration of the effects of AD on the semantic network. Specifically , the salience of a particular dimension for categorizing conceptsand individual differencesin thesemantic network could not be examinedwith the data from the verbal fluency task becauseof the variance in the total number of responsesproduced by each . To circumvent these deficiencies , we conducted a second study subject to measure the proximity between concepts using a more systematicprocedure , et al., 1993b , Salmon ). , Butters (Chan In this secondstudy, the proximity (i.e., strength of association ) between in which a task triadic was estimated subjectschose by comparison concepts alike. A sample two that are most the from among three concepts(i.e., animals ) of twelve high frequency animal names served as the concepts in this task and every combination of three animal nameswas presentedfor a total of 220 trials. This procedure produced a proximity score reflecting the strength of associationfor eachpair of animalsin relation to all of the other animal names ; that is, how frequently those two animals were chosen as most alike. The proximity scoreswere subjectedto an MDS analysiswhich . Inaddition produced cognitive maps for AD patients and elderly NC subjects to the cognitive mapsfor the two groups, the proximity data were used to calculatefor eachindividual subject a quantitative measureof the relative importance of the various dimensionsfor categorizing concepts(i.e., dimension weights) and a measureof the degreeof relianceon one or more of the index). dimensions(i.e., skewness The cognitive maps of AD and NC subjects revealed with this analysis were best representedby three dimensions that appearedto correspond to ), predation (herbivore vs. carnivore), and domesticity (i.e., wild vs. domestic size (large vs. small) (figure 16.2). Although a three-dimensional solution provided the best spatial representationof the semanticnetwork of both AD and NC subjects , severalsignificant alterations were evident in the semantic . network of the AD patients
, patientswith Nq subjects Figure 16.1 The cognitivemapsof elderly normalcontrol (E T) obtained of the Alzheimer with dementia ' s disease (HD), andpatients type ( DA Huntington a verbal data from on and from multidimensional fluency analyses scaling clustering performed in the animal name is determined ; animals task . Thepositionof each by multidimensional scaling .) et al., 1993a fromChan . (Adapted areencircled same cluster together

385

' : SemanticNetwork Abnormalities in Alzheimer s Disease I: Butters Chan ,& , Salmon

t
t

lop

.'I .~ .) :i

ltd
lop
. .
a

.
.

DOn

flQq8

MOO as
Ita

e ..1 iN I

ilain J
.

-. J

lop

DOg
as
.

3f

ai

Mm

I ..a

alJl8
.

Ilqq
Juwqdata

Ja8

11Qq8

M03

tqf

First, AD patients focused primarily on concrete conceptual information i.e. , whereascontrol subjectsstressedabstract ( , size ) in categorizing animals , a number of animals that conceptualknowledge (i.e., domesticity). Second were highly associatedand clustered together for NC subjects were not strongly associatedfor patients with AD . Third , patients with AD had significantly lower skewnessindices than the NC subjects , indicating that they were less consistent in utilizing the various attributes of the animals (predation . It should be noted that , domesticity, and size ) for categorization theseresultsare not likely to be a consequence of random responding by the AD patients (due to a lack of attention or an inablity to follow instructions), becausevirtually identical results were obtained when the patients were tested again with the sameproceduresfive to sevendays later. The semanticnetworks of the AD patients and NC subjectswere further , explored by performing Pathfinder analyses (Dearholt 8t Schvaneveldt 1990) on the triadic comparison task data generated from the Chan et al. , Butters , Salmon , et al., 1995). The Pathfinderanalysisis (1993b ) study (Chan more effective than MDS for examining the number and strength of the associationsconstituting the semanticnetwork. This analysis systematically constructsa model of the semanticnetwork that consistsof nodes representing . It , and links representing the connections between concepts concepts has been proposed that personswho possesssufficient knowledge in a particular domain to evaluate the degree of associationbetween concepts will , develop a clear and concisenetwork composedonly of relevant connections whereasthose who never developed this degreeof knowledge, or who have lost knowledge they once possessed , will tend to organize conceptsin a relatively chaotic way with many unnecessary . When the adjunct connections -generatedcognitive mapsof the AD patients were comparedwith Pathfinder those of the NC subjects , they were characterizedby abnormal strength of associationbetween usually weakly related conceptsand a large number of adjunct links (figure 16.3). unnecessary In contrast to the cognitive maps generated for the AD patients , the semantic network models generated by MDS and Pathfinder analyses for HD patients and patients with a circumscribedamnestic syndrome did not -matchedcontrol subjectsin terms differ significantly from those of their age of the number of dimensionsemployed, the strength of associations , or the number of common links between concepts . These results provide further evidence that factors such as general cognitive decline , deficiencies in retrieval process es , or impaired episodic memory have little impact on the nature of the cognitive map generatedby thesemultidimensionalanalysesof data from semanticmemory tasks . Thus, the distorted cognitive map of the AD patients most likely reflects semanticmemory disruption that may arise from extensivedamageto the associationcortices (McCarthy 8t Warrington , 1988; 1990; Warrington 8t Shallice , 1984).

387

: SemanticNetwork Abnormalities in Alzheimer : Butters ' s Disease Chan , Salmon .&

'S DETERIORA nON OFTHESEMANTIC I ME R NETWORK OF ALZHE DISEASE PATIENTS OVERTIME


While the studies reviewed above provide compelling evidence that the structure of semanticknowledge is disrupted in patients with AD , little is known about how semanticnetworks are altered as the diseaseprogress es. Because AD is a progressive degenerativedisorder in which the neocortical associationareasbecome increasingly degraded (Braak 81 : Braak , 1991), we ' that AD hypothesized patients semanticnetworks would becomemore and more disorganizedover time. To explore this hypothesis , we recently examined the relationship between semantic network deterioration and the : Salmon , Butters severity of global dementia in patients with AD (Chan , 81 , 1997), assumingthat the severity of global cognitive impairment is a reflection of the degreeof pathological abnormality. In this study, a standard semantic network for animals was generated , based on triadic comparison task data , for a group of elderly NC subjects . A semanticnetwork generatedfor eachAD patient using an MDS analysis -of -fit measure was then evaluatedagainst this standardnetwork and a goodness was calculatedbasedon the percentageof variancein the AD patient' s semantic network that could be accounted for by the standard network. -of-fit measurewas related to severity of dementia as When the goodness measuredby the Dementia Rating Scale (DRS ; Mattis , 1976), a significant negative correlation was observed indicating that the integrity of the semanticnetwork decreases as dementiaprogress es. Furthermore , qualitative assessmentof the results of the MDS analysis demonstrated that AD patients tended to rely less on abstract attributes (e.g., domesticity), and more on perceptual attributes (e.g., size ), in categorizing animals as they becamemore demented , and the primary dimension used to classify animals on the triadic comparisontask becamelesssalient as the severity of dementia increased . An examination of the relationship between dementia severity and the semanticnetwork model generatedwith the Pathfinderanalysisin this study indicated that the complexity of the network and the strength of association between concepts changes with disease progression . The complexity of the semanticnetwork was evaluatedby comparing the number of links in the standard semantic network of a group of elderly NC subjects with the number in group networks generated for mildly , moderately, and severely -generatednetworks of all three AD dementedAD patients. The Pathfinder groups containedmore links than that of the NC subjects(but the difference between the NC and mildly dementedAD subjectsonly approachedsignificance ) and the AD groups differed significantly from each other, with more links being evident as the severity of dementia increased . This increasein the complexity of the semanticnetwork with increasingdementia was confirmed by a correlational analysiswhich demonstrateda significant negative

389

Chan t Butters : Semantic NetworkAbnonnalities in Alzheimer , Salmon ,& ' s Disease

correlation between the number of links in each AD patient' s network and his or her scoreon the DRS. To further explore the relationship between the integrity of the semantic network and the severity of dementia :x : for , we calculated a similarity inde eachAD patient basedon the number of links that his or her Pathfindernetwork had in common with the standard control network. The higher the similarity index for a given patient, the more intact the semanticnetwork. A correlation analysisrevealedthat the integrity of the semanticnetwork (i.e., the similarity score ) was correlated with the level of dementia as measured by the DRS scoresof the AD patients. This finding suggeststhat the AD ' patients understandingof the relative associationamong conceptsis altered throughout the course of the diseaseand the organization of their semantic es. knowledge becomesmore and more abnormalas the disease progress Another recent study demonstrated that the integrity of the semantic network can predict the subsequentrate of global cognitive decline in AD , Butters , et al., 1995). A highly significant correlation , Salmon patients (Chan ' was found betweenAD patients similarity indicesgeneratedfrom Pathfinder , and their subsequent analysesas describedabove global cognitive decline as measuredby changesover one year in their scoreson the DRS (figure 16.4). Given that semanticknowledge may be stored in a distributed fashion within

, r-squared= .84 y = 61.852x- 38.145


5

Jeal

~ ~ ! 0 C /) ~ Q . N

2. 5 0 . 2. ~ .5 . 7.5 . 10 . 12. 5 - 15 - 17 .5 - 20 0 . 22. 5 .3 .35 .4 .45 .5 .55 .8 .65 .7 0 0 0 0 0

.75

Semantic

Network

Similarity

Index

's disease with Alzheimer networksimilarityindexof twelvepatients Figure 16.4 The semantic between the of rateof cognitive decline asmeasured by the difference plottedasa function Dementia obtained nearthe timeof semantic ) scores (DRS knowledge testing(year RatingScale .) et aI., 1995 1) andapproximately oneyearlater(year2). (Adapted from Chan

390

in Dementia Applications

the associationcortices ( Warrington & Shall ice, 1984), this result suggests that early evidence of semantic memory dysfunction in AD patients may serve as a marker for the susceptibility of these cortical regions to further deterioration and a resulting increasein the severity of global dementia .

CONCLUSIONS
The results of the studies reviewed in this chapter indicate that network models of the organization of semantic knowledge are abnormal in patients with AD . While the source of this abnormality remains somewhat controversial : Johnson, 1995 ), the models may reflect a deterioration (Bonilla & in the organization of semantic memory , or an actual loss of semantic knowledge , in patients with AD . This degradation of semantic memory may arise from the gradual deterioration of the neocortical association areas in which semantic knowledge is thought to be stored (McCarthy & : Warring ton , 1988 , 1990 ). It should be noted , however , that semantic memory is not totally destroyed in AD . Even in the later stages of their disease, theseman tic networks of AD patients remain grossly intact with interpretable dimensions and some normal associative links between concepts . It is the change in the number and quality of these associations, and changes in the salience of the dimensions characterizing concepts , that differentiate them from NC subjects . The normal semantic network models generated for HD and amnestic ' patients suggest that the disorganization seen in the AD patients networks are not due to nonspecific factors such as global cognitive impairment , episodic memory impairment , or a deficiency in retrieval processes. The patients with HD that took part in the studies described above were as severely demented as the AD patients in most cases , and a general retrieval deficit has been proposed as a primary characteristic of their dementia syndrome (see Heindel , Salmon, & : Butters , 1993 ). Similarly , the amnestic patients ' episodic memory deficits were as severe as those of the AD patients . It is also important to note that of these three patient groups , only AD patients develop semantic memory dysfunction , and only AD patients develop significant neuropathological changes in the cortical association areas.

ACKNOWLEDGMENTS
Dr. Butters passedaway before the publication of this book. We dedicate this chapter to his memory. The preparation of this chapter was supported in part by NIA grant AG -O5131and NIMH grant MH -48819 to the University of California at SanDiego.

REFERENCES
Anderson architecture . Cambridge , J. R. (1983 Press . ). The , MA: Harvard of cognition University : Tomoeda , K. A., & in dementia , C. K. (1983 . Brain ). Confrontation Bayies namingimpairment andlAnguage . , 19 , 98- 114

391

: Butters : Semantic NetworkAbnormalities Chan , Salmon in Alzheimer ,& 's Disease

Bonilla, J. L., & : Johnson . (1995). Semanticspacein Alzheimer ' s diseasepatients , M. K . Neuropsychology , 9, 345- 353. : Braak Braak , H., & , E. (1991). Neuropathological staging of Alzheimer-related changes . Ada , 82, 239- 259. Neuropathologica Butters , N., Granholm : Wolfe, J. (1987 , E., SalmonD ., Grant, I., & ) . Episodic and semantic . Journalof Clinical and Erptrimtntal memory: A comparison of amnesicand demented patients , 9, 479- 497. Neuropsychology Chan , A . 5., Butters ., & , N., Paulsen : Maloney, L. T. (1993a , J. 5., SalmonD . P., Swenson , M. R ). An assessment of the semanticnetwork in patients with Alzheimer ' s disease . Journalof Cogni Hve Neuroscience , 5, 254- 261. : SalmonD . P. (1997 Chan , A 5., Butters , N., & ) . The deterioration of semantic networks in -sectionalstudy. Neuropsychologia 's disease : A cross , 35, 241- 248. patients with Alzheimer Chan , A 5., Butters , N., SalmonD . P., Johnson : Swenson . (1995). , S. A ., Paulsen , J. 5., & , M. R . Neuropsychology , 9, Comparisonof the semanticnetworks in patients with dementiaand amnesia 177- 186. : McGuire, K. A . (1993b Chan , A . 5., Butters , N., SalmonD . P., & ). Dimensionality and clustering in the semanticnetwork of patients with Alzheimer ' s disease . Psychology and Aging, 8, 411419. : Johnson Chan , A 5., SalmonD . P., Butters , N., & , S. A (1995). Semanticnetwork abnormality 's disease . Journalof the predicts rate of cognitive decline in patients with probable Alzheimer Intern aHonalNeuropsychological , 1, 297- 303. Society Chertkow, H., & : Bub, D. (1990). Semanticmemory loss in dementiaof Alzheimer ' s type. Brain , 113 , 397- 417. Collins, A . M ., & : Loftus, E. F. (1975). A spreadingactivation theory of semanticprocessing . Psychological Review , 82, 407- 428. Collins, AM ., & : Quillian, M . R . (1969). Retrieval time from semanticmemory. Journalof Verbal and VerbalBehavior , 8, 240- 247. Learning Dearholt, D. W., & : Schvaneveldt . W. , R. W. (1990). Properties of Pathfinder networks. In R Schvaneveldt(Ed .), Pathfinder associa Hve networks : Studiesin knowledge Hon (pp. 1- 30). organiza Norwood , NJ: Ablex. Fuster cortu : An empiricalapproach , J. M . (1995). Memory in the cerebral to neuralnetworksin the humanand nonhuman . Cambridge . , MA : MIT Press primate Heindel, W. C., SalmonD . P., & : Butters es to the memory disorders , N. (1993). Cognitive approach of dementedpatients . In P. B. Sutker & : H. E. Adams ( Eds .), Comprehensive handbook of . , 2nd ed., (pp. 735- 761). New York: PlenumPress psychopathology : Patterson , J. R., & , K. (1995). Is semanticmemory consistently impaired early in the Hodges course of Alzheimer ' s disease ?: Neuroanatomicaland diagnostic implications. Neuropsychologia , 33, 441- 459. : Butters , J. R., SalmonD . P., & , N. (1990). Differential impairment of semanticand episodic Hodges ' ' : A controlled prospective study. Journal memory in Alzheimer s and Huntington s diseases and Psychiatry , Neurosurgery , 53, 1089- 1095. of Neurology . , SalmonD . P., & : Butters , J. R , N. (1991). The nature of the naming deficit in AlzheimHodges er's and Huntington' s disease . Brain , 114 , 1547- 1558. . , SalmonD . P., & : Butters , J. R 's , N . (1992). Semanticmemory impairment in Alzheimer Hodges disease : Failureof access or degradedknowledge ? Neuropsychologia , 30, 301- 314.

392

in Dementia Applications

Mattis , S. (1976 ). Mentalstatus examination for organic mental in the elderlypatient . syndrome In L. Bellack & : T. Karasu .), Geriatric . New York (Eds : Grune & : Stratton . psychiatry . A,& : Warrington -specific ,R , E. K. (1988 for modality McCarthy ). Evidence in systems meaning thebrain . Nature - 430 . , 334 , 428 : Warrington , R. A., & , E. K. (1990 McCarthy of semantics ). Thedissolution . Nature . , 343 , 599 : Sdtvaneveldt , D. E., & . W. (1975 Meyer ,R ). Meaning structure es . , andmental , memory process In C. N. Cofer(Ed .), The structure : Freeman Francisco of human (pp. 54- 89 . ). San memory . J., & , L J., Shoben : Smith ,E , E. E. (1973 distance andthe verification Rips ). Semantic of semantic relations . Journal andVerbal Behavior of Verbal , 12 , 1- 20. Learning , A. K., Shepard : Nerlove , R. N., & Romney , S. B. (1972 ). Multidimensional and scaling : Theory in thebehavioral sciences : Seminar , Vol. 2. New York applications Press . Rumelhart : McClelland , D., & : the PDPResearch , J., & distributed ). Parallel Group(1986 processing : in themicrostructure . 1: Foundations . Cambridge &plorations , Vol : MIT Press of cognition . , MA Shallice , T. (1988 to mental ). From structure . Cambridge . : Cambridge neuropsychology , U. K University Press . , R. N., Romney : Nerlove , A K., & , S. B. (1972 Shepard ). Multidimensional and sailing : Theory in thebehavioral sciences , Vol. 1. NewYork: Seminar applications Press . : Katzman , R. D., & . (1983 . R dementia of the Alzheimer Terry ). Senile . Annals , of Neurology type - 506 14 . , 497 , R. D., Masliah , E., SalmonD. P., Butters Terry . , Hill, R , N., DeTeresa . , Hansen , R : , L. A , & Katzman , R. (1991 basis of cognitive ). Physical alterations in Alzheimer ' s disease : Synapse lossis themajorcorrelate - 580 of cognitive . Annals . , 30 , 572 of Neurology impainnent : Hutchinson , A., & , J., W. (1986 Tversky ). Nearest of psychological . neighboranalysis spaces Review , 93 , 3- 22. Psychological . K., & : shallice ,E , T. (1984 ). Categoryspecific semantic Warrington . Brain , 107 , impairments 829 - 854 .

393

: Butters : SemanticNetwork Chan , Salmon ,&

in Alzheimer ' s Disease

17

Parallel Distributed in Models Processing


Alzheimer ' s Disease Lynette J. Tippett and Martha J. Farah

Much of the researchon cognitive impairments in Alzheimer' s disease(AD ) has been aimed at identifying the underlying locus or loci of processing . Cognitive neuropsychologists impairmentswithin the cognitive architecture commonly use two types of experimentalstrategy to localize the underlying . The first impairment in persons with AD or other forms of brain damage involves the analysisof error types made on tasks . For example if , the errors in a naming task are visual, in the sensethat the nameproduced belongs to a different but visually similar object, then researchers infer that the locus of processing impairment is visual. Similarly, semantic errors suggest breakdowns at semanticstages , and so on. The secondmain approachto inferring the locus of processing impairment is the selective manipulation of the . For example , if the difficulty of difficulty of individual task components lexical retrieval in a naming task is increasedby selecting objects with low ' , andthis manipulationdisproportionately affectsAD patients frequencynames , then researchersinfer that the locus of processing naming performance is . impairment lexical Alternatively , if a person is disproportionately affected , this suggestsan impairment in by reducing the visual quality of a stimulus the visual stages of processing . These strategies provide straightforward meansof investigating functional impairmentsin brain-damagedpersonsand . drawing conclusionsfrom the results In this chapter we explore alternative ways of interpreting the performance of personswith AD , making use of the conceptsand tools of parallel distributed processing(PDP) (Rumelhart : the PDP Research , McClelland, & 1986 . The PDP can lead to ) very different conclusionsconcerning Group, approach the underlying loci of impairment in persons with brain damage , 1994). In the first part of the chapter (Farah , we review the results of researchaimed at identifying the locus of damage responsiblefor the confrontation naming deficit in AD . According to the traditional types of reasoning , the effects of selective manipulations of difficulty in just described confrontation naming have seemed to implicate both visual and lexical impairments in AD . We then demonstrateby simulation that these results are consistent with an impairment confined to semanticmemory, assuming that are distributed and interactive. In the secondpart of the representations

chapter we provide an alternative interpretation for naming errors that seem, according to traditional means of neuropsychological inference , to imply that semantic knowledge of categories and exemplars is separate and that AD primarily affects the latter . Again using a simulation of confrontation naming , we show how an undifferentiated semantic memory store would , when damaged, lead to preserved knowledge of categories because of the greater degree of initial learning for knowledge of categories in distributed representations .

NAMINGIMPAIRMENTS IN ALZHE I ME R'S DISEASE : SEMANTIC , VISUAL LOCUS OF DAMAGE ? , OR LEXICAL


The first question we addressed with a PDP approach emerged from attempts to localize the functional locus of impairment of the naming difficulty in AD . Difficulty in naming objects is an early and consistent sign associatedwith AD , which typically worsens as the disease progress es. the vast amount of data in collected studies with AD , Despite participants the functional deficit underlying the naming problem is not well understood . One explanation is that it results from impaired semanticmemory. According to this view personswith AD cannot name objects becausetheir semantic . One sourceof evidence memory knowledge about them is impoverished for this comes from the qualitative nature of errors made in naming tasks . Typical kinds of errors include calling an object either by the name of its superordinatecategory (e.g ., animal instead of cat) or by the name of a semantically relatedobject (e.g ., lettuce for asparagus : Tomoeda , ) (e.g ., Bayies& 1983; Hodges : Butters 1991 Martin & : Fedio . 1983 These errors , Salmon ,& , ; , ) are degradedor underspe suggestthat the availablesemanticrepresentations cified. In addition to evidencefrom naming tasks themselves , evidencefrom other experimental paradigms also implicates semantic memory as a major locus of cognitive impairments in AD (e.g., Chan, Butters & : Salmon , 1997; : Hodges , Patterson , Graham , et al., 1996; Ralph , Patterson & , 1997; Hodges also seeNebes , 1989 , 1992, for reviews). As a result of both sourcesof evidence , impaired semanticmemory is currently the predominant explanation for the naming deficit in AD . Nevertheless , there remain two alternative accounts , each of which ' appearsto be supported by direct evidence from AD subjects performance . One account is basedon the finding that AD subjectsare on naming tasks disproportionately affectedby reduction of the visual quality of the stimulus to be named (e.g ., line drawings vs. objects : Kelly, ; e.g ., Kirshner, Webb, & : Huber, 1988), suggestinga deficit of visual perception. 1984; Shuttleworth & The second account arises from the finding that AD subjects are differentially impaired at naming objects whose namesare low -frequency words, relative to objects whose namesoccur with high frequency (e.g ., Barker & : Lawson : Jones : Huber, , 1968; Skelton Robinson & , 1984; Shuttleworth &

396

Applications in Dementia

1988), which suggests a deficit of lexical access . Furthermore , phonemic ' : cuing has been reported to improve AD subjects naming (e.g ., Martin & Fedio , 1983; Neils, Brennan , Cole, et al., 1988), again implicating lexical or phonological retrieval as the locus of impairment. This range of results appearsto implicate three different loci of impairment , visual, and lexical underlying the naming deficit in AD , namely, semantic . This multiloci account may be valid, either as a result of heterogeneity of impairments between different persons with AD , or becauseindividual subjects may have multiple loci of impairment. An alternative view, however , is that a semanticmemory impairment alone is sufficient to account for the data directly implicating semanticmemory as the locus of impairment, and for the apparently conflicting findings implicating visual perceptualand lexical factors. This alternative view rests on two key ideas &om PDP: distributed and interactivity . In a highly interactive and distributed representation , different parts, or components , of the system are in systemof representation a constant give-and-take relationship , continuously updating eachother with the partial, intermediate results of their processing . Each part of an item' s , and lexical) dependson other parts for representation(e.g ., visual, semantic , so that damageconfined to one part of the system input to becomeactivated will render all parts less able to attain their proper activation states . In such a system of that are to , manipulations difficulty targeted nondamaged parts of the system would have an increasedeffect, becausethe functioning of the nondamagedparts is neverthelesscompromised . Thus, if AD patients have a deficit of semanticmemory, reducing the quality of the visual input may affect naming performancemore than in persons with intact semantic memory becausetheir visual systems are deprived of some of the normal . Similarly, producing words that occur with low support &om semantics &equency may be more vulnerable to damage to semantic memory than . Providing additional producing words that occur with high &equency activation to the lexical or phonological componentsin the form of phono, regardless of logical cuing may facilitate activation of the correct name whether there is damagein the lexical component . Theseideashave their roots in the last century, when neurologists pointed out the interactive nature of brain organization , and the resultant possibility that widespreadchangesin functioning could follow a localized lesion (e.g ., Ferrier , 1886). Somemodern neurologists (e.g ., Kinsbourne , 1977) have also an interactive of neural information . championed conception processing has raised the of a 1980 close interaction ) Goodglass( specifically possibility " between semanticsand lexical access , suggesting that inadequateactivation of a concept prevents retrieval of its name ." Caramazza , Berndt, and Brownell (1982) and Bayies and Tomoeda (1983) have also emphasizedthe interaction between strictly perceptual and strictly lexical process es in con&ontation naming. Hinton and Shallice(1991) and Plaut and Shallice(1993) have documented the effects of interactivity between vision and semantics on errors in reading , confirming that error type is not always transparently

397

's Disease : POP Modelsin Alzheimer Tippett& Farah

. What has been lacking, however diagnostic of the underlying locus of damage , are studiesthat explicitly investigate thesegeneralideasof interactivity , semantic , and lexical components of confrontation among perceptual in to order naming interpret the effects of visual and lexical difficulty on naming performancein AD . We attempted to do just this by implementing a PDP model of confrontation naming in AD . " " , we trained an interactive neural network to produce name Specifically " " patterns when presentedwith visual patterns by way of an intermediate " " set of semantic , thus simulating confrontation naming. We then patterns " " investigated the claim that damageto the semantic component will result in enhancedvulnerability to manipulations affecting nondamaged components " " , specifically the quality of visual inputs, frequency of training of : Farah , and effectsof phonological cuing (Tippett & , 1994). patterns

The Model

In POP models , representationsconsist of a pattern of activation over a set of highly interconnectedneuron-like units. The extent to which the activation of one unit causesan increaseor decrease in the activation of a neighboring unit depends on the " weight" of the connection between them; positive weights causeunits to excite eachother and negative weights cause units to inhibit eachother. For the network to learn that a certain name representation , the weights goes with a certain visual stimulus representation in units the network are so that adjusted among presentation of either the visual pattern in the visual units or the namepattern in the nameunits causes the corresponding other pattern to becomeactivated. Upon presentation of the input pattern to the input units, all of the units connected with those input units will begin to change their activation under the influence of two kinds of constraints : the activation value of the units to which they are connected . Theseunits might in turn connect and the weights of the connections to others, and influencetheir activation levels in the sameway. In recurrent , or attractor, networks, such as the model to be presented here , the units downstream will also begin to influence the activation levels of the earlier units. Eventually, these shifting activation levels acrossthe units of the network settle into a stable pattern, or attractor state. The attractor state into which a network settles is determinedjointly by the input pattern (stimulus ) and the weights of the network (stored knowledge). Figure 17.1 shows the architecture of the model. There are five pools of " units. The 16 visual units" subservevisual representationsof objects. The " " 32 semanticunits subserverepresentations of the semanticknowledge that " " " " can be activated either by visual inputs, or by name inputs. The 16 " name units" subserve the representationof namesassociatedwith objects. The model is trained to associate the patterns of activity between thesethree layers. The learning of these associationsis assistedby the presenceof the " " , Hinton , & Williams, remaining two unit pools, the hidden units (Rumelhart

398

Applications

in Dementia

SE M A Nn C UNfI 'S (32 )

Figure17.1 Functionalarchitectureof model of confrontation naming.

1986). Theseare located between the visual and semanticlayers (the " visual " hidden units ) and the name and semanticlayers (the " name hidden units" ). All units within eachlayer are connectedto eachother, and all connections , both between and within layers , are bidirectional. Thus, there is a high degree of interconnectivity in the model, which allows for extensive interaction among the different componentsin the course of naming. In addition to weights between units, there are bias weights for eachunit, which encode the relative likelihood of eachunit being activated over all trained patterns. Twenty patterns consisting of patterns in the visual layer, semanticlayer, and namelayer, were generated . Visual representations and namesof objects are representedby randomly generatedpatterns of activity (- I ' s and + 1' s) over the 16 units in each pool . Similarly, the semanticknowledge activated by theseinputs is representedby a random pattern of activity over the pool of 32 units in the semanticlayer. The representationsare distributed, in the sensethat an object is representedby a pattern of activation over visual, semantic , and nameunits, and in the sensethat within eachof these pools of units all of the units participate in representing all of the known patterns. " Eachunit encodesinformation that representssome " microfeature ( Hinton, McClelland, & : Rumelhart , 1986), but there is no attempt to assign units to specific easily labeled microfeatures (e.g ., " red" or " round" for a visual ). Of course , there is nothing intrinsically visual, semantic , representation or verbal about these pools of units, aside from the fact that their patterns of mutual connectivity conform to the general notion that semantic

399

: Farah : POP Models in Alzheimer ' s Disease Tippett &

must be accessed in order to mediatebetweenvisual and name representations . representations Units in this model can take on continuous activation values between - 1 and + 1. The weights on the connectionsbetween units can take on any real . Activation levels are updated according to a nonlinear activation values function (a hyperbolic tangent). The network was trained to correctly associate the patterns of activity between the three layers , such that when an was to one of the e. . the visual layer), it was layers ( g , input pattern presented able to produce the correct patterns of activity at the other two layers (the semanticlayer and the namelayer). During training, as well as testing, inputs were soft-clamped , that is, their activation values were not fixed, but were the result of input activation along with activation from other units to which . Noise was injected into eachlayer to facilitate robustness they were connected of learning. To assess the generality of the results , multiple networks were trained with the same patterns but with different random starting weights. The learning procedure used during training was the contrastive Hebbian learning (CHL) algorithm (Movellan, 1990). Note that we are using CHL as a meansof setting the weights in the network to enableit to perform naming; we are not concernedwith the psychological reality of the training procedureas a model of learning. Confrontation naming was simulatedby presentingjust the visual pattern to the visual input layer, allowing the network to settle, then looking at the pattern of activation on the nameunits to seewhich of the 20 namepatterns it matchedmost closely. The degreeof pattern match was simply the number of units whose values(+ 1 and - 1) matched . It was therefore , a forced choice between the twenty patterns representingnames . Damage to semantic memory was simulated by removing randomly chosen subsetsof the semanticunits. Four levels of damage were used , to the effects of severe : 2 4 units units 6 units , , , explore increasingly damage and 8 units, or 6.25 percent , 12.5 percent , 18.75 percent and 25 percent of the semanticlayer. Although the neuropathologicalchangesin AD are more , we chose this means of lesioning the complex than a simple loss of cells network because it was the simplest and most straightforward. Furthermore , Patterson , Seidenberg , and McClelland (1989) compared the effects on network performanceof deleting units, deleting weights, and adding noise , and found them similar. Not all changes to a network necessarily have : In their neural network simulation of AD , Chan and comparable effects associates Marshall Chan , , Butters , et al., 1992) found distinctive effects of ( . Nevertheless , the issuesof interest selectively lesioning inhibitory connections in this study could be addressed . by this more generic form of damage
: in Alzheimer ' s Disease Simulationsof Naming DeAcit

The first question of interest was the claim that if the components underlying naming are highly interactive, damage to semanticscould render the

400

Applications in Dementia

system more sensitive to manipulations of the visual quality of the stimulus . This was investigated by testing trained undamaged and damaged networks on the production of the correct name patterns with normal visual inputs and impoverished visual inputs . Five undamaged trained networks were used. Twenty different random lesionings at each level of damage were then carried out for each network . Visual impoverishment of the input patterns was simulated by reducing the inputs of a randomly chosen 50 percent of the units in the input pattern . For example , if part of the pattern was + 1 + 1 - 1 - 1, then the impoverished input might be + .5 + 1 1 - .5. Two levels of visual impoverishment were used for both damaged and undamaged networks . The first was used so that the naming task was as difficult as possible for the intact network without resulting in any errors . The second level of visual impoverishment used was selected so that the intact network made at least one error . These levels were motivated by the observation that normal subjects make few or no errors in naming impoverished stimuli . While it was expected that damage to the semantic layer would impair network performance on the naming task overall , the question of interest was whether the damaged networks would show enhanced vulnerability to impoverished visual inputs relative to the intact networks . The results showed that although the amount of interaction between visual impoverishment and damage becomes smaller at increasing levels of damage, there is a highly consistent and significant interaction at all levels for the first level of impoverishment (figure 17.2) and for sixteen out of twenty of the damaged networks at the second level of impoverishment (figure 17.3 ) (see Tippett & : Farah, 1994 , for further details ). In other words , the effect of visual degradation was greater on networks with semantic damage than on intact networks . In our simulations , normal performance is at or near ceiling . This raises the possibility that the appearance of increased sensitivity to visual impoverishment in the damaged networks is due to a ceiling effect in the intact networks . Of course, the same point can be made about the empirical data from AD subjects and normal controls : normal subjects invariably perform at or near 100 percent in confrontation naming , even with the presentation of impoverished visual stimuli (e.g ., see data of Kirshner et al., 1984 ). We do not think the high performance of normal subjects undermines the logic of the experiments using visual impoverishment , for the following reason. These studies were cast within the traditional framework of assumptions of discrete stage models (e.g ., Sternberg , 1969 ) in which processors (such as a visual processor ) transfer the results of their processing to further stages only after successful completion of that stage of processing . Using these assumptions , if it were true that AD patients have normal perception , then they should be just as much at ceiling in the visual processing of stimuli as normal controls . Therefore , with respect to those manipulations of visual processing difficulty for which normal subjects were at ceiling , AD patients should be just as immune , by virtue of the same ceiling effect. If this were the

401

's Disease : PDPModelsin Alzheimer & I: Farah TippeH

p8 Indul

.I I c~ 1 ,I I -i I=
.

' slndu

'

P8 ~

g ~ai ~g0
~~. jjO ~

~ 0 (

0 (

Ian

I ~. JJO~

IO8JJOO

Applications in Dementia

0809Ot 001

p Indul

Ot Ot oz 0 001 08
-

p Indu

8 . - i i 8~c
J~. JJO~ pa % " ( P84 .

0 ~0 G0 .. 0 N0 -0
~ % I ~. JJO

P8481 ~
f . .. f -f . f ~

~ 1 . ft ~ -

t
. P84 0 (

403

: Farah : PDP Models in Alzheimer ' s Disease Tippett &

09

08

case , there would be no interaction causedby visual impoverishment , and ' AD patients generally lower performance on all confrontation naming would be due to problems downstream , suchas semanticmemory. Of course , if a systemwere highly interactive, suchlogic would not apply. As we demonstrated in this simulation in a highly interactive network, ) at one level can interact manipulations of difficulty (visual impoverishment with impairment of a component at another level (semanticdamage ), while these same manipulations of difficulty have little or no effect on networks with intact semantics . The second question of interest is the prediction that a naming system with highly interactive componentswill have more difficulty producing the names of items that have occurred with lower frequency during learning than those that occurredwith higher frequency , after damageto the semantic . This is based on the claim that in an interactive system , items component trained with lower frequencywill be lessrobustly represented , and therefore more vulnerable to damage in any component of the system . The investigation was carried out by training a seriesof twelve networks with stimulus items presentedat different frequencies , that is, ten of the stimulus items were presentedat a lower frequency than the other ten items. Six networks were trained with an item-frequency ratio of 4 : 1 and six networks with an item frequency ratio of 3 : 1. Networks were then damaged by removing randomly selectedunits in the semanticlayer and were tested on the full set of items used in training. Each network was subjected to ten different random lesionsat eachof four levels of damage . As predicted , more errors were made in the production of namesfor items trained at low frequency , compared with items trained at high frequencyfor networks trained on each of the two ratios and at each level of semantic damage from 6.25 percent to 25 percent (see figure 17.4 for results of networks trained at the 4 : 1 ratio). With no damage , all the networks produce all namesfor the low - and high-frequency items without error. With damagein the semanticlayer, networks make differentially more errors producing namesfor the low -frequencyitems. Thesefindings were significant on twenty two out of twenty -four tests that compared the effects of the four levels of semanticdamageon naming the high frequency items versus the low frequency items in networks trained at the ratio of 4 : 1. For the six networks that were trained at the ratio of 3 : I , sixteen out of twenty -four tests were significant . These data support the hypothesis that lower-frequency items will suffer disproportionately after damageto semanticsin an interactive naming system . However, in this simulation the frequency manipulation involved different amounts of training for the visual and semantic , as well as the lexical, levels of representation . In effect, our lexical frequency manipulation was confounded with other types of frequency . However, the sameis undoubtedly true, on average for the concrete nouns used in confrontation naming , tasks , even though the degree of correlation among lexical, semantic , and

404

in Dementia Applications

GOL 08 Ot

"

.c D %
IO8JJOO %

E . 1~8JJO~

Aouenbe

~ A

q (

001 08 09 ot

.c Ut X

405

' s Disease : PDPModelsin Alzheimer : Farah Tippett&

OO 08 O

0n nCD

" Table 17.1 Effectiveness of "phonemic with damage in the semantic units : cuing on networks Meanpercentage of correct name on failed items responses previously Levelof D~
-

6.25 %
of 10 Sample lesioned networks M = 67.6% SD= 32.40 range 0% /0 - 1000 0 1 2 2 (/ / )

12 .5%
M = 81.5% SD= 12.15 range 600 /0 - 100% 3 5 ( / 7/ 7)

18 .75 %
M = 73.5% SD= 15.93 range 50010 - 100% (3/6- 9/ 9)

25 %
M = 71.2% SD= 7.24 range 58' 0- 83.3 % - 10 (7/ 12 ) / 12

visual frequency would be less perfect in reality than it was in our simulation . In fact, without controlling for somenonlexical measureof familiarity, the lexical frequencymanipulation cannot be used to implicate a specifically -stage architecture lexical locus of impairment . For this , even in a discrete reason the results of with AD , phonemic cuing patients seema more decisive form of evidencefavoring a lexical locus of impairment. If a phonemic cue helps a patient produce the nameof an object, it is generally inferred that the locus of naming impairment is lexical, as opposed to semantic . The final manipulation applied to this model was to test the validity of this inference . Will the naming performanceof an interactive system , with damageconfined to the semanticlayer, be facilitated by presentationof " a partial " phonemic pattern, just as the naming of objects by AD patients is facilitated by the provision of a phonemic cue ? A random sample of ten networks from the first simulation was tested , using only those damaged items on which they madeerrors. For eachsuch item, the network was given a " phonemiccue." This was accomplished by soft-clamping 25 percent of the name units with the correct name pattern while the visual pattern was presented " " to the visual units, thus simulating, for example , saying br to facilitate the naming of a picture of a broom. For networks at all levels of damage , phonemic cuing frequently enabled correct naming performanceon those items that had previously been failed. Table 17.1 provides a summary of the mean percentage of correct name responsesproduced following cuing. The efficacy of cuing ranged from a meanof 67.6 percent for networks with 6.25 percent damageto the semantic units, to 81.5 percent for networks with 12.5 percent damage . At the highest level of damage to the semanticunits (25 percent the mean improvement ) was 71.24 percent . These findings demonstrate that an unambiguously lexical or phonological manipulation is effective in a semanticallydamaged . system In summary , these simulations were undertaken to attempt to reconcile apparently conflicting data in the literature about the underlying impairment of the naming deficit in AD . In an interactive PDP model, it was possible to demonstratethat if one componentof the systemis damaged , in this casethe

406

in Dementia Applications

semanticcomponent , the remaining componentswere impaired in their functioning if , only to the extent of being more easily taxed by difficulty manipulations such as visual impoverishment . Similarly, when processing in one was bolstered additional external by component input, as in phonemic cuing, this helped a systemwhose damagewas in another component . The findings are consistentwith the view that a single locus of impairment in AD , namely in semanticmemory, is sufficientto accountfor a wide variety of data related to naming tasks , including those thought to implicate other loci of impairment . Of course , this does not imply that there is always only one locus of in AD . Schwartzand Stark (1990) have reviewed a large body of impairment evidence on the clinical and pathological course of AD , and conclude that there is considerablevariability in the systemsaffected , and in the stage of illness at which they are affected . Thesefindings clearly do not support the general strategy of inferring the locus of cognitive impairment on the basis of sensitivity t~ manipulations which selectively affect the processing of one component at a time. The model demonstratesthat this approach is not reliable in highly interactive , since both visual and phonemic manipulations yielded positive systems . This suggests results in our model when the damagewas confined to semantics ' that the analysis of patients cognitive impairments will have to make use of more fine-grained measuressuch as distributions of error types, and , perhaps in conjunction converging operations among experimental tasks with computationalmodeling.
Preservation of Categorical Knowledge in Alzheimer ' s Disease

with the use of neural network models A secondquestion that we addressed relatesto the nature of the semanticmemory impairment in AD . A common observation is that in persons with AD , semanticknowledge of exemplars often appearsdisproportionately impaired relative to knowledge of superordinate . The dissociation between exemplar and category categories of semanticmemory, including knowledge can be seen in various measures " " " " confrontation naming (e.g., naming a desk as furniture or a chair ) category and attribute verification, picture-name matching, and category fluency : Seidenberg e. . : T omoeda , 1989; Hodges , 1983; Chertkow, Bub, & ( g , Bayies& : Growdon, 1986; Martin , et al., 1991; Hodges , et al., 1996; Huff, Corkin, & : Fedio, 1983; Ober, Dronkers, Koss 1987; Martin & , et al., 1986; Troster, : Trosset Salmon , 1996 for a , McCullough, et al., 1989 but seeCox, Bayies& counter finding). The most transparentinterpretation of theseobservationsis that the brain honors the distinction between different levels of a hierarchical , categorical AD and that affects the neural , systems subserving knowledge system primarily PDP . The of rise to a second give principles exemplar knowledge interpretation of the dissociation found in AD , however, which does not involve separatesystemsfor representingthese two kinds of knowledge. In

407

's Disease t Farah : PDPModelsin Alzheimer Tippett&

POP systems , knowledge representationis widely distributed, with a manyto -many mapping of things representedand things (units) doing the representing . In other words, many units are involved in a representation , and individual units are involved in many representations . Becauseof this, semanticmemory may be conceptualizedas a distributed system of knowledge , in which the same physical substrate representsboth representation categorical and exemplar information. The relative preservation of category of the greater robustnessof the category attributes knowledge is a consequence (i.e., those attributes sharedby all or most membersof a category) compared -unique attributes. This to exemplar greater robustness is a direct of the increasedfrequency of learning of category attributes consequence . In other words, the resulting from the more inclusive nature of categories attributes of a semantic are learned and activated each category pattern time a member of that category is presented or encountered . In contrast, -unique attributes become active only when that exemplar particular exemplar is presented . Simulations of Preserved Category Knowledge in Alzheimer ' s Disease We tested and confirmed the computational adequacyof this hypothesis in two computer simulations (Tippett , McAuliffe & Farah , 1995). We trained the sameinteractive network describedabove (see figure 17.1), to associate " " " " " " patterns in pools of visual units with patterns in semantic and lexical units. As before, there is nothing intrinsically visual, semantic , or verbal about these pools of units, aside from the fact that their patterns of mutual connectivity conform to the general notion that semanticrepresentationsmust be accessed in order to mediatebetween visual and namerepresentations . In these simulations the twenty patterns in the semanticlayer were constructed so that four categorieswere represented , each with five individual . Half of each semantic that is , 16 of the 32 units of the exemplars pattern, semanticrepresentation , representedthe properties sharedacrossexemplars within a category. This portion of the pattern was therefore identical across the five exemplars . The other 16 units representedthe individual exemplar for each item and overlapped with other exemplars within the knowledge 8 of the 16 units. An example is provided in figure 17.5, category by only which shows the semanticpatterns of the five membersof categories 1 and 2. The other categories were constructed in a similar manner except that they shareddifferent setsof 16 units. Specifically , the category portion of the semanticpatterns for category 1 involved semanticunits 1 though 16, category 2 involved semanticunits 9 through 24, category 3 involved semantic units 17 through 32, and category 4 involved semanticunits 1 through 8 and 24 through 31. Thus, all semantic units functioned equally often to representcategory and exemplar knowledge. There is no division of labor, within the semanticlayer, for category vs. exemplarknowledge.

408

in Dementia Applications

CAT EGO RYl


SEMANTIC UNITS 1. 8
+ - + - + - + + - + - + - + + - + - + - + + - + - + - + + - + - + - + -

9-16
+ - + - + - + + - + - + - + + - + - + - + + - + - + - + + - + - + - + -

-24 17
+ + + + + + + + - - - - - - - + + + + - - - - - - - + + + + + - + - + - + -

.32 25
+ + + + + + + + + + + + + + + + + + + + - - - + + + + - - - - + - + - + - +

CATEGORY 2
SEMANTIC UNITS 1- 8 ++++++++ - - - - - - - ++++- - - - - - - ++++ +- +- +- +9 - 16 ++++++++++++++++++++17 - 24 ++++++++++++++++++++25 - 32 ++++++++ ++++++++ ++++- - - ++++- - - - +- +- +- +

Figure 17 .5 Semantic patterns of the five members of category 1 and the five members of category 2. Note that units 1- 16 carry the category infonnation for category 1, and units 9- 24 carry the category infonnation for category 2.

Ten networks were trained to associatetwenty patterns of activity on the visual, naming, and semantic layers. Each network was trained with the semanticpatterns describedabove , but the visual and naming patterns were different for eachof the ten networks. As before, damageto semanticmemory was simulated by removing randomly chosen subsetsof the semantic units. Four levels of damagewere used to explore the effects of increasingly severe damage : 2 units, 4 units, 6 units, and 8 units, or 6.25 percent , 12.5 18.75 and 25 of the semantic . At each of these , , percent layer percent percent levels of damage , each network was lesioned in twenty different ways, that is, with twenty different random patterns of semanticunits eliminated . The results of this simulation are summarizedin table 17.2. After damage , the system showed a stronger tendency to within -category errors than -category errors, as do personswith AD , even though the between damaged representationsdo not separatelyimplement category and exemplar knowledge . While appearing to confirm our hypothesis as to the greater robustness of category-level knowledge to damage , in fact the reason for these results is ambiguous . It is well known that similar patterns may be confused with one another after network damage(e.g., Hinton & : Shallice , 1991; Plaut & : Shall ice, 1993). Therefore , it is possiblethat this finding simply reflectsthe greater similarity among exemplarswithin a category than among exemplars of different categories , rather than being a result of the greater robustnessof the category level knowledge.

409

's Disease : Farah : POPModelsin Alzheimer Tippett&

Table 17.2 Average number of multiple- choicenaming errors out of twenty , at four levels of , showing expectedand observedwithin- category errors damage Level of Damageto Semantic Layer (%) 6.25 12.50 18.15 Expected Within Category Errors 0.53 1.11 1.68 Observed Within Category Errors 0.19 1.16 2.39 Significance Level (Binomial Test) 0.001 0.001 0.001

Average Overall Errors 2.5 5.58 1.91

25.00

10 .64

2.24

2.91

0.001

Table 17.3 Average number of erroneousunit activation values per pattern in category and

at four levels of damage exemplar portionsof semantic patterns

6.25 12.50 18.75 25.00

0.20 0.48 0.80 1.04

0.24 0.59 0.88 1.12

0.001 0.001 0.001 0.001

In order to check this possibility we carried out a second simulation to assess directly the state of category and exemplarknowledge after semantic . Instead of looking at the names produced by the damaged layer damage network, we looked at the semanticlayer itself and compared the relative accuracyof the category vs. exemplar portions of the patterns. Table 17.3 shows the averageerror in category and exemplarportions of the patterns in ten different networks after twenty different random lesionings at eachlevel of damage . There is a consistenttendency for the categoricalinformation to -unique information, even be more robust to damage than the exemplar though errors in category and exemplar units should be equally likely by chance . In other words, in this simulation the preservationof categoricaland -unique semanticknowledge is assessed directly, without the confounding exemplar influence of the greater number of similar and hence confusable . patterns within a category than between categories on our error measure The results support the claim that categorical information is robust to damage becauseof the greater degree of training enjoyed by categorical information. This simulation illustrates that observed neuropsychologicaldissociations do not invariably and transparently reflect the structure of the underlying , 1994; Shallice , 1988), particularly in the caseof cognitive architecture(Farah in dissociations which two abilities , may rely on the samearchitectural single but one , may be more resistant to damage than the other component

410

Applications in Dementia

, 1955). In the caseof the dissociation between category and exemplar ( Teuber knowledge observed in AD , it is not necessaryto hypothesize that semanticmemory is compartmentalizedinto category-level knowledge and , our explanation is an instanceof exemplarknowledge. In very general terms ice (1988), in Teuber both noted the type of explanation (1955) and Shall by " robust" than " " which the relatively impaired ability is simply easier or more " " " " the relatively sparedability . Easy and robust are not mechanisticterms, . The however; rather they are descriptive labelsin an incomplete explanation these rather tests it is that the directly advantageof computational approach , and in so doing provides an explicit, mechanistic nonspecific concepts . accountof the phenomena In thesesimulations , there are two featuresof neural network computation which are particularly relevant to the explanation of preserved category knowledge in AD , namely, distributed representationand graded learning. With distributed representationsof knowledge, the category is implicitly of the category by the attributes sharedby all or most members represented ' s of each . The 17.5 see ) ( exemplar representationis categoricalportion figure more therefore encountered . than the exemplar frequently, during learning networks unique portions. Becauselearning is not all or none in connectionist , , but a matter of degree which depends on, among other factors in a trained overlearned is more of , frequency training category knowledge network than exemplarknowledge.

NEUROPSYCHOLOGY nO NS FORCLINICAL
Do findings from these POP models have implications for the practice of ? Indeed they do. In our view , these findings have clinical neuropsychology both specific implications for understanding cognitive impairment in AD , of cognitive impairments and general implications for the clinical assessment . of brain with various of individuals damage types of cognitive impairment in personseither Neuropsychologicalassessment suspectedof, or with , a diagnosisof probable or possible AD is challenging of the variability in the cortical regions affectedand in the stage of because illness at which they are affected , and the consequentvariability in cognitive deficits. An increase in knowledge about impairments most consistently , such as semantic present during the early to middle stages of the disease tests in clinician assist the will , selecting appropriate for assessing memory The . of areas critical findings from our POP models cognitive functioning semantic memory is a locus of support other research suggesting that impairment in most patients with AD . In addition, they have shown that in an interactive system a semanticimpairment alone can account for patterns of performancethat might at first glance appear to indicate another type of cognitive deficit (e.g ., the sensitivity of AD patients to phonemic cuing on ). naming tasks

411

~1sDisease Modelsin A17Mim : POP : Farah Tippett&

The brain has long been viewed as a complex interactive system by many academicand clinical neuropsychologists . Nevertheless , applying this general principle in a meaningful way to our understandingof the consequences of damage to specific components of the cognitive architecture has been very difficult, and insteadinterpretation of patient performancehas tended to be much more literal. Sensitivity to visual factors in tasks , or the production of visual errors, has , for example , been interpreted as indicating impairment in the visual stage of processing . In other words, for want of any clear alternative there has been a tendency to interpret impaired neuropsychologic , specific performanceas if brain and cognition operateas discretestage , rather than as interactive systems systems . The advent of connectionist , however, provides an alternative: POP models provide theory and POP models a meansto reasonabout the complex relations between interactive systems and behavior and the effects of damaging such a system . Importantly the approachalso allows us to test in explicit mechanisticterms suchreasoning . Thus, for example , the POP models described in this chapter have shown that differential sensitivity to manipulations of visual factors on naming tasks does not reliably indicate a deficit at the visual stage of processing in highly interactive systems . The relation between results of such manipulationsand the locus of impairment in the damagedcognitive system is much lesstransparent . Such findings clearly have implications for interpreting cognitive performance ' in brain-damaged individuals generally. Interpretation of patients cognitive impairmentsmust make use of more fine-grained measures such as distributions of error types and converging evidencefrom a variety of tasks . Many clinical neuropsychologistsare well practicedat looking at patterns of , so the adaptation required may be less one of procedural performanceacrosstasks , and more one of the reasoningabout, and interpretation of, change . performance

WHERE TO FROMHERE ?

We have used two examplesto illustrate the application of computational , to questions emerging from the neuropsychologica modeling, in this casePDP models literature on cognitive impairments in AD . These simulations have allowed us to test, in an explicit and mechanisticmanner , generalclaims madeabout the properties of interactive systems . One of the major criticisms of computational models , however, is that they are post hoc; they account for old data without making new contributions or predictions. For this reason we believe it is important that novel predictions emerging from these models be tested with AD subjects . This is important both from the point of view of checkingthe validity of the computationalmodel and from our need to extend further our knowledge and understanding of the impairments experiencedby personswith AD . In this way the relation between computa tional modeling and empirical work with personswith brain assumes damage

412

Applications in Dementia

: both componentsin a constant givethe properties of an interactive system and-take relation, continually updating each other with the partial, intermediate . results of their processing A CKN 0 WL EDGMENTS Committee was supportedby a University of Auckland Research This research ' Association Pilot Research Grant Alzheimer s L. . T. and awarded to J , grant MH48274 93 IO621 , PRG 93 153, ONR grant NOOO14 , NIMH grant RO1 in Cognitive Scienceat the and NSF STC grant to the Institute for Research , University of PennsylvaniaResearchResearch University of Pennsylvania M . to awarded Foundation , J. F.

REFEREN CFS
. British in dementia , 114 , : Lawson Barker , J. (1968 }oumalof Psychiatry , M., & ). Nominalaphasia - 1356 . 1351 in dementia . Brain : Tomoeda , C. K. (1983 ). Confrontation , K. A., & namingimpairment Bayies . andLanguage , 19 , 98- 114 : Perceptual delict hypothesis : Brownell , H. H. (1982 ). The semantic Caramazza , A., Berndt , R. 5., & - 189 15 161 . and . Brain , classification , and Language by object patients aphasic parsing in patients networks of semantic . (1997 : SalmonD ) . Thedeterioration , N., & Chan , A. 5., Butters - 248 . . Neuropsychologia : A cross -sectional , 241 's disease , 35 with Alzheimer study in semantic : SalmonD. (1992 ). Undmtanding , N., & memory Chan , A. 5., MarshallS., Butters 's disease annual convention at the fourth . Presented ~ ch A : Alzheimer appr patients computational . 1992 of theAmerican , SanDiego , June Sodety Psychological : Seidenberg Chertkow ). Primingand semantic , D., & memoryloss in Alz, H., Bub , M. (1989 ' . - 446 andLanguage . Brain s disease heimer , 36 , 420 deterioration andattributeKnowledge : Trosset Cox , M. W. (1996 ). Category , K. A., & , D. M., Bayies - 550 . andLanguage . Brain 's disease in Alzheimer , 52 , 536 andBrain . Behavioral brain with an interactive inference Farah , M. J. (1994 ). Neuropsychological 104 . 17 43 Sciences , , . . London : Smith , Elder Ferrier , D., (1886 ). The of thebrain functions ' canScientist of naming , 68 , 647 , H. (1980 ). Disorders followingbraininjury. Amm Goodglass . 655 . In D. : Rumelhart , D. E. (1986 Hinton ). Distributed , J. L., & , G. E., McClelland representations distributed .), Parallel : the PDPResearch Rumelhart ,& : processing , J. L. McClelland Group (Eds - 109 . , MA: MIT Press in themicrostructure , Vol. 1 (pp. 77 ). Cambridge of cognition Erplorations of acquired : Investigations an attrador network : Shallice Hinton , T. (1991 ). Lesioning , G. E., & 98 74 95. Review . , , dyslexiaPsychological . (1996 : Dawson . R., Patterson ,K , N., & ). Namingandknowingindemen , K., Graham ,J Hodges - 325 . andLanguage 's type . Brain , 54 , 302 tia of Alzheimer : Butters , N. (1991 ). The natureof the namingdelict in Alz, J. R., SalmonD. P., & Hodges ' ' . . Brain s disease s andHuntington , 1547 1558 heimer , 114

413

's Disease : POPModelsin Alzheimer : Farah Tippett&

Martin , A. (1987 of semantic ) . Representation and spatialknowledge in Alzheimer ' s patients : for modelsof preserved . Journal Implications learningin amnesia of Clinicaland Experimental - 224 . , 9, 191 Neuropsychology Martin : Fedio ,A,& , P. (1983 andcomprehension ). Word production in Alzheimer ' s disease : The - 141 breakdown of semantic . Brain and . , 19 , 124 knowledge lAnguage Move Uan , J. (1990 Hebbian ). Contrastive . In D. S . learningin the continuous Hop6eldmodel : T. J. Sejnowski , G. E. Hinton ,& .), Proceedings Touretzky (Eds models of the 1989connectionist summer school - 17 Mateo (pp. 10 ) . San . , CA: MorganKaufman Nebes . D. (1989 ,R ). Semantic ' s disease . Psychological Bulletin memoryin Alzheimer , 106 , 377 394 . Nebes . D. (1992 ,R in Alzheimer ). Cognitive ' s disease . In F. I. M. Craik& : T. A Salt dysfunction - 446 house .), The handbook andcognition (Eds : Erlbaum of aging (pp. 373 . ). Hillsdale , NJ Neils , J., Brennan , M. M., Cole Uer : Gerdeman , M., Bo , F., & , B. (1988 ). Theuseof phonemic cuing - 354 with Alzheimer 's disease . Neuropsychologia . , 26 , 351 patients

Ober , B. A , Dronkers , N. F., Koss : Friedland , E., Delis , D. C., & , R. P. (1986 from ). Retrieval -type dementia semantic . Journalof Clinicaland Experimental memory in Alzheimer Neuropsychology - 92. , 8, 75 Patterson , K , Seidenberg : McCle Uand , J. L. (1989 Mectionsand disco , M. S., & ). Co Mections : in a computational modelof reading dyslexia es . . G. M. Morris (Ed Acquired .), Parallel process In R distributed - 181 : Implications and neurobiology processing for psychology : (pp. 131 ). Oxford Oxford University . Press Plaut : Shallice , D. C., & , T. (1993 andsemantic ). Perseverative influences on visualobjectnaming errorsin opticaphasia : A cO Mectionistaccount . Journal Neuroscience . , 5, 89- 117 of Cognitive , M. A. L., Patterson : Hodges . (1997 , K., & , J. R Ralph between ) . The relationship namingand semantic for differentcategories in dementia of Alzheimer ' s type . Neuropsychologia knowledge , - 1260 35 . , 1251 Rumelhart , D. E., Hinton : Williams , G. E., & , R. J. (1986 internalrepresentations ). Learning by errorpropagation . In D. Rumelhart , J. L McClelland : the POPResearch ,& .), Parallel Group(Eds distributed : Erplorations in themicrostructure .1 : Foundations processing , Vol of cognition (pp. 318 362 . ). Cambridge , MA: MIT Press Rumelhart Uand , D., McCle : thePOPResearch , J. L , & distributed : ). Parallel Group(1986 processing in themicrostructure . 1: Foundations . Cambridge Erplorations , Vol of cognition : MIT Press . , MA Schwartz : Stark , M. F., & , J. A. (1990 ). Oinicopathological models of Alzheimer 's disease and seniledementia : Unravelingthe contradictions . In M. F. Schwartz . ), Modulardeficits in (Ed -type Alzheimer dementia (pp. 61- 82). Cambridge . , MA: MIT Press Shallice , T. (1988 to mental ). Fromneuropsychology structure . Cambridge , U. K : Cambridge University Press .

414

in Dementia Applications

. of Alzheimer of dementia disorder : Huber . C., & worth , S. J. (1988 ). Thenaming type Shuttle ,E . .1 .2- 234 Brm ' n - lAng ,1 ~ , 34 dementia -Robinson andthe severityof senile : Jones , S. (1984 ). Nominaldysphasia Skelton . M., & 171 . 145 168 . British , , Giry of Psychi ] 0NrnQ1 . Ada ' s method of Dander : Extension of processing . (1969 ). The discovery stages Sternberg , S . 315 30 276 , , J Psychologil - 296 . Rtoiew . Annual , 6, 267 H. Teuber of Psychology ). Physiological , L (1955 psymology : ' s disease modelof namingin Alzheimer : Farah , M. J. (1994 ). A computational , L J., & Tippett . , 8, 3- 13 Neuropsychology Unitaryor multipleimpainnents1 in of categorical : Farah , M. J. (1995 ). Preservation knowledge , S., & , L. J., McAuliffe Tippett . - 533 . Memory account , 3, 519 : A computational Alzheimer ' s Disease of category : Butters , N. (1989 ). A comparison Tr6ster , A L, SalmonD. P., McCullough , D., & 's disease Brain . , , 37 ' s and Alzheimer with associated Language Huntington 8uencyde6dts . 500- 513

415

@ l '5Disease inAI7: Mim Models : FarahPOP Tippett6

Index

models deficits Acalculia andprefrontal - 249 calculations , 246 , 233 - 235 functions , 233 cognitive - 246 andmatching memory , 245 - 252 in neuropsychology , 251 - 238 number , 235 processing - 240 numerical , 238 quantities , 241 246 numerosity - 251 Wisconsin Test CardSorting , 249 , , 52 , 146 , 170 (A Ch ), 34, 37 Acetylcholine - 323 321 - 323 ACh , 170 , 321 , 34, 37, 52, 146 retrieval for Actions , lexical - 287 , 286 background , 289 findings functional , 289 (PET ) studies imaging - 289 lesionstudies , 287 , new Activememory , 84, 96 , 48- 51 Activeneurons , 225 Activerepresentation , 359 's disease AD. See Alzheimer theory(ART) Adaptiveresonance to traditional asalternative syrnbol 4 13 15 20 26 models , , , , processing rule and 106 , pseudonormalization Wisconsin CardSortingTestmodeland , - 362 361 - 265 ADHD, 69, 264 , 267 -drivenhelplessness Affect , 150 Akineticmutism , 67- 68 Alcoholdependence " desire "abnormal to drinkand , 132 , 134 and 8 129 networks , , backpropagation - 137 model , 125 computational 132 136 desire to drink , of alcohol , 136 dependence development - 133 of model , 125 development - 135 of model results , 133 - 137 treatment of alcohol , 136 dependence

genetic wlnerability and, 132 134 neural circuitry and, 124- 125 and, 123- 124 neurobiologicalmechanism 124 and neurotransmitters , " normal" desireto drink and 132- 134 , - compulsivedisorder and, 125obsessive 126, 128, 132, 136 research , 123- 125 serotonergicfunction and, 123 124 therapeuticmodalities and, 133 135 Alexia, 237 238 Algorithms , 11- 12, 128- 129, 218 backpropagation BESA , 66 contrastive Hebbianlearning , 400 ' Alzheimer s disease (AD ) breakdown in network function and, 320321 categoricalknowledge in, preservationof, 407- 411 cognitive impairment in, 395 methodology and, 12 gradient-descent humanmemory model and, 320- 321 , 371implications of, neuropsychological 373 languagedeficit and, 382 naming impairmentsin, 396 398, 400- 407 of, neuroanatomical aspects 370- 371 neurobiologicalaspectsof, 370- 371 , 396parallel distributed processingmodels 413 architecture , 398- 400 , 407- 411 categoricalknowledge preserved 412 413 future directions , , 411implications for neuropsychology 412 lexical impairments , 398- 400 semanticimpairments , 396- 398, 400- 407 visual impairments , 398- 400

. of Alzheimer of dementia disorder : Huber . C., & worth , S. J. (1988 ). Thenaming type Shuttle ,E . .1 .2- 234 Brm ' n - lAng ,1 ~ , 34 dementia -Robinson andthe severityof senile : Jones , S. (1984 ). Nominaldysphasia Skelton . M., & 171 . 145 168 . British , , Giry of Psychi ] 0NrnQ1 . Ada ' s method of Dander : Extension of processing . (1969 ). The discovery stages Sternberg , S . 315 30 276 , , J Psychologil - 296 . Rtoiew . Annual , 6, 267 H. Teuber of Psychology ). Physiological , L (1955 psymology : ' s disease modelof namingin Alzheimer : Farah , M. J. (1994 ). A computational , L J., & Tippett . , 8, 3- 13 Neuropsychology Unitaryor multipleimpainnents1 in of categorical : Farah , M. J. (1995 ). Preservation knowledge , S., & , L. J., McAuliffe Tippett . - 533 . Memory account , 3, 519 : A computational Alzheimer ' s Disease of category : Butters , N. (1989 ). A comparison Tr6ster , A L, SalmonD. P., McCullough , D., & 's disease Brain . , , 37 ' s and Alzheimer with associated Language Huntington 8uencyde6dts . 500- 513

415

@ l '5Disease inAI7: Mim Models : FarahPOP Tippett6

Index

models deficits Acalculia andprefrontal - 249 calculations , 246 , 233 - 235 functions , 233 cognitive - 246 andmatching memory , 245 - 252 in neuropsychology , 251 - 238 number , 235 processing - 240 numerical , 238 quantities , 241 246 numerosity - 251 Wisconsin Test CardSorting , 249 , , 52 , 146 , 170 (A Ch ), 34, 37 Acetylcholine - 323 321 - 323 ACh , 170 , 321 , 34, 37, 52, 146 retrieval for Actions , lexical - 287 , 286 background , 289 findings functional , 289 (PET ) studies imaging - 289 lesionstudies , 287 , new Activememory , 84, 96 , 48- 51 Activeneurons , 225 Activerepresentation , 359 's disease AD. See Alzheimer theory(ART) Adaptiveresonance to traditional asalternative syrnbol 4 13 15 20 26 models , , , , processing rule and 106 , pseudonormalization Wisconsin CardSortingTestmodeland , - 362 361 - 265 ADHD, 69, 264 , 267 -drivenhelplessness Affect , 150 Akineticmutism , 67- 68 Alcoholdependence " desire "abnormal to drinkand , 132 , 134 and 8 129 networks , , backpropagation - 137 model , 125 computational 132 136 desire to drink , of alcohol , 136 dependence development - 133 of model , 125 development - 135 of model results , 133 - 137 treatment of alcohol , 136 dependence

genetic wlnerability and, 132 134 neural circuitry and, 124- 125 and, 123- 124 neurobiologicalmechanism 124 and neurotransmitters , " normal" desireto drink and 132- 134 , - compulsivedisorder and, 125obsessive 126, 128, 132, 136 research , 123- 125 serotonergicfunction and, 123 124 therapeuticmodalities and, 133 135 Alexia, 237 238 Algorithms , 11- 12, 128- 129, 218 backpropagation BESA , 66 contrastive Hebbianlearning , 400 ' Alzheimer s disease (AD ) breakdown in network function and, 320321 categoricalknowledge in, preservationof, 407- 411 cognitive impairment in, 395 methodology and, 12 gradient-descent humanmemory model and, 320- 321 , 371implications of, neuropsychological 373 languagedeficit and, 382 naming impairmentsin, 396 398, 400- 407 of, neuroanatomical aspects 370- 371 neurobiologicalaspectsof, 370- 371 , 396parallel distributed processingmodels 413 architecture , 398- 400 , 407- 411 categoricalknowledge preserved 412 413 future directions , , 411implications for neuropsychology 412 lexical impairments , 398- 400 semanticimpairments , 396- 398, 400- 407 visual impairments , 398- 400

Alzheimer ' s disease .) (AD) (cont - 391 semantic networkmodel , 382 - 391 deterioration of, 389 - 382 functional view of, 381 studies of 383 388 , past - 373 verbalAuency modeland , 369 " American Psychological Society modelin " 153 154 , principle - 180 modulation - 177 , 176 , 179 Aminergic Amnesia - 78 , 77 , 80- 82, 300 anterograde term and , 86- 87 long memory - 78 , 77 , 82 retrograde -termmemory short and . 86- 89 , 64- 65 Amplification , spatially distributed , Amplitude gradient 106 - 107 . 37, 46- 47 Amygdala of variance (ANDVA) measures Analysis , 129 , 131 Animalsystems . See also , 19 models specific ANDVA measures , 129 , 131 - 78 amnesia . 77 , 80- 82 , 300 Anterograde Arousal 343 345 , system , ART. See Adaptiveresonance theory , 84 Articulatory loop Artificialneural networks . See also , 8- 10 , 334 specifictypt S - 360 Associative learning , 359 - 303 Associative , 301 memory Attention . See also disorders specific activememory and . 49 and . 64- 65 amplification - 72 controland . 71 cognitive disorders 72 , 69, frontallobeand . 48- 51 historical of, 61 - 62 perspective and James , 61 neural networks and . 62- 71 executive , 65- 68, 70 interaction of, 69- 71 orienting , 62- 65, 69- 70 , 68- 70 vigilance lobe and . 62- 63 parietal , 341 spatial colliculus and , 63- 64 superior Attentiondeficithyperactivity disorder - 265 (ADHD), 69 , 264 , 267 Attentional controller , 84 Attentional disorders attention deficithyperactivity disorder , 69 , - 265 264 , 267

networks and ,8 backpropagation - 268 model 261 computational attention - defidthyperactivity disorder - 265 simulation , 264 - 264 simulation 1, 263 , 2S hemineglect - 261 lettercancellation simulation . 265 - 268 limitations of, 261 - 261 linebisection simulation , 265 - 263 normal attentional es , 261 , 261 process 266 261 extinction . explaining , 259 , 259 sensory simulation of, 266 - 261 - 261 , 251 neglect anatomical - 261 network , 260 - 259 assessing , 251 - 261 explaining , 259 distributed models , 260 parallel processing - 263 Attentional es , normal261 , process 261 Attentional tasks , 65- 68 of primates , 36- 31, 41 Auditory system Autoassodative function of recurrent collaterals in CA3 , 308 Automatic , 19 , 83- 86 , 94 workingmemory ISO 152 , , Automaticity Autonomic nervous . 37 system - 12 - 129 , 11 , 128 , Backpropagation algorithm 218 networks Backpropagation alcohol and , 8, 129 dependence asalternative to traditional symbol models , 4, 7- 12 , 26 processing attentional disorders and ,8 - 332 basal modeland , 331 ganglia with complex human neuropsychological - 12 test , 11 and ,8 neglect , 8 Strooptestand - 12 Towerof Hanoiand , 4- 5, 11 understanding , 8- 10 Basal ganglia neural networks and , 70 , 331 in orientingnetwork , 70 ' s disease Parkinson anddysfunction of, 333 roleof, 331
structure of, 331 Basalganglia model , 331- 332 backpropagationnetworks and , 340- 341 cognitive networks and future directions , 347- 349

418

Index

- 347 limbicnetworks and , 342 - 343 based contingency learning , 342 's disease in Parkinson and , 346 depression 347 - 346 disorders , 343 neuropsychiatric structure 332 , , 342 - 340 motornetworks and , 332 - 335 motorcontrol , 334 - 340 motorlearning , 338 motormaps 337 338 , Parkinson disease and , 333 - 337 mechanisms , 335 pharmacologic structure , 332 333 roleof basal and , 331 ganglia structure of basal and , 331 ganglia Basic objectlevel 273 Behavioral states . See specific types BESA , 66 algorithm Bias nodes , 361 , 363 Bias units , 131 Bilateral extensive lesions , 87 , 90- 91 of, 23 , problem Biological plausibility 165 , rhythm Biological Brainelectrical source (BESA ) analysis , 66 algorithm Brainimaging . Setalso specific types electrical , 93 - 96 resonance , 93 magnetic emission , 93- 96 positron tomography Brainorganization , vertebrate , 36- 37 Brainstem andthalamic models , 173 Calculations - 249 , 233 , 246 interference of, 23, problem Catastrophic 25 Catecholamine , 34, 52, 112 - 15 network , 13 Categorization 363 field . Cateogory Central executive attentional controller , 84 Central nervous . Set also (CNS ), 195 system specific components CHLalgorithm , 400 , 311 Cholinergic degeneration - 311 modulation , 119 , 309 Cholinergic neurons , 110 Cholinoceptive - 12 - 262 , 66- 68, 11 , 261 Cingulate Circadian , 165 rhythm "Circuits " level 234 251 of circuits , , Clinicalresearch , 25- 26. Setalso specific types , 383 analysis Clustering . See CNS also , 195 specific components CNV paradigm , 91- 98, 104

Cognition acalculiaand prefrontal deficits models , 233235 cortex and, 37- 40 developmentof, 52- 53 , 51- 52 energy requirementsand human , 33, 61 processingcapabilitiesand, 51- 52 waking and sleepingstatesand, 165- 168 Cognitive control, 71- 72 , 142, 144- 145 Cognitive depression , 383, 385 Cognitive maps Cognitive networks basalganglia model and, 340- 341 in primates , 36- 37, 47 auditory system and brain organization and , vertebrate , 36- 37 , 33 cognition and, human developmentof, 52- 53 frontal lobe systemsand , 48- 51 future models , 53 54 nervous system of primatesand, 33- 34 neuronsand , 34- 35, 42- 43 neurotransmittersystemsand, 34- 35 , 35 olfactory system and , 33, 35 parallel distributed processingand relations betweenprocessingcapabilities and energy requirements , 51- 52 somatosensory systemand, 36- 37, 47- 48 visual systemand, 36- 37, 40- 47 , 299 Cognitive neuroscience Cognitive psychology, 299 Cognitive systemsinterpretation of , 152- 153 depression es , 109- 110 Cognitive-motor process , 359 Competitive selection . Setalsospecific types Computational models alcohol dependence , U5 - 137 desireto drink , 132- 136 , 136 developmentof alcohol dependence developmentof model U5 133 resultsof model 133- 135 treatment of alcohol dependence , 136- 137 attentional disorders , 261- 268 attention-defidt hyperactivity disorder simulation , 264- 265 , 257, 263- 264 heminegled simulation letter cancellationsimulation , 265- 267 limitations of, 267- 268 line bisection simulation , 265- 267 normal attentional process es , 261- 263, 267 cellular level, 8- 10

419

Index

. See alsospecific Computationalmodels types (cont.) human memory, 299- 300 , 172- 177 waking and sleepingstates brainstemand thalamicmodels , 173 distributed and cortical models , 173- 177 overview, 172 results , 182 ) , 277 Computerizedtomography (CT , 273 Conceptualknowledge Concreteentities defined , 273 , 273- 282 nonunique , 273- 275 background defined 273 , findings, 281- 282

functional - 281 (PET , 280 ) studies imaging - 280 lesions studies , new , 275 - 286 , 282 unique - 283 , 282 background defined , 273 - 286 , 285 findings functional (PET , 285 ) studies imaging - 285 lesions studies , new , 283 - 110 , timingandclassical , 109 Conditioning Connectionist models modelof freeandcuedrecall hippocampal - 306 and , 305 - 192 and , 191 , 206 schizophrenia of Sloman andRumelhart , 371 termof, 3, 8- 10 -scheduling Contention mechanism , 65 Context - dependent lexical , disambiguation - 206 203 Contextual , 273 complexity - 343 -based Contingency learning , 342 variation (CNV ) Contingent negative , 97- 98 , 104 paradigm Continuous firing ratemodel - 318 freerecallin, 317 - 317 associate in, 315 paired learning Continuous test , ), 193 performance (CPT - 203 201 , 264 , 267 Contrastive Hebbian (CHL ) learning , 400 algorithm - 147 Control versus , mastery , 145 Correlational learning , 81 . See Cortex also Prefrontal cortex(PFC ) and , 37- 40 cognition entorhinal , 309 rhinal , 50 Cortical . See Cortico - hippocampal processing andmemory system

Cortico - hippocampal andmemory system - 78 amnesia and , 77 , 80- 82 - 78 , 77 , 80- 82 , 300 anterograde 77 78 , , 82 retrograde brainimaging , 92- 96 electrical , 93 resonance , 93- 96 magnetic - 96 emission , 93 positron tomography - 112 in, 111 dimensions functional model , 98- 103 and , 77 , 80 learning link- operator storeand , 78 - 80, 112 -tennmemory - 113 and , 78 long mathematical model , 104 multirange bufferand , 78 dataand , 86- 92 neuropsychological rehearsal allowed , 86- 89 - 92 rehearsal , 89 prevented -tennmemory short and , 78 , 80, 107 , 112 113 - 108 ordernetworkmodels , 104 temporal astemporal , 79 80 process - 111 timing , 109 - 80, 83- 86, 96and , 78 workingmemory 98 - 203 CPT , 193 , 201 , 264 , 267 CT, 277 Cuedrecall . See modelof free Hippocampal andcuedrecall
DA . See Dopamine Darwinian scheme , 235 Dedsion rule deAdts, 152

Declarative , 77 , 81- 82 memory of retention 89 , Delay -to- sample match task , 49 Delayed -response tasks , 249 Delayed - 390 Dementia Scale ( DRS ), 389 Rating Dentate 306 308 , gyrus . See also Learned and Depression helplessness model depression computational and , 150 , 152 automaticity 142 144 145 , , cognitive of, 152 systems cognitive interpretation 153 - 143 motoric , 142 -mediated , 150 , 152 norepinephrine 's disease in Parkinson - 347 , 346 and , 150 , 152 perseveration of, 133 persistence - 144 , 142 vegetative states and , 172 wakingandsleeping on freerecall - 320 , 319 Diazepam

420

Index

Dissociations , 20- 21 , 274 Distributed andcortical networkmodels , - 177 173 Distributed , 397 representation DLPFC , 357 , 365 A) (D Dopamine interactions with otherneurotransmitters , 345 internal controland , lossof, 147 learned and , 150 helplessness - 198 and , 195 schizophrenia - 196 effects of, 195 neuromodulatory effects of 197 198 , physiological - 196 cortexand , 195 prefrontal Dorsal , 42 pathway - 366 Dorsalprefrontal cortexmodel , 365 Dorsolateral cortex DLPFC ( ), 357 , prefrontal 365 - 390 DRS , 389 . See Drugs specific types -task -recall Dual , free , 101 paradigm

Facial , 42 recognition Fast ruleswitching , 249 Feature field , 363 Feature , 337 map Feedback , 145 loops 5-HT, 34, 52, 146 , 170 , 345 fMRI, 61, 66, 92 , 228
Forebraindemodulationmodel. 176- 177 Freerecall . 91- 92. 317- 320. See alsoHippocampalmodel of &ee and cued recall Frontal lobe. 48- 51 alsoFrontal lobe . 196. See function model; Tower of Hanoi (TOH ) test. in &ontaliobe -lesionedpatients Frontal lobe function model dorsal versusorbital pre&ontal cortex model and. 365- 366 prindples of neural networks and. 359- 361 theoretical modeling of. 357- 359 verbal Ruencymodel. 366- 369 Alzheimers disease and. 369- 373 Wisconsin Card Sorting Test and. 357- 358 Wisconsin Card Sorting Test model and. 361- 364 Frontal-lesionedpatients also . 248- 249. See Tower of Hanoi ( TOH) test. in &ontal lobe-lesionedpatients Functionalimaging (PET ) studies for actions . 289 for nonuniqueconcreteentities. 280- 281 for unique concreteentities. 285 Functionalmagneticresonance imaging (fMRI ). 61. 66. 92. 228 Functionalmodel of hippocarnpo -cortical relations . 98- 103

- 315 EEG , 92, 167 , 171 , 228 , 313 Efferent control 40 , Electrical brainimaging also , 93. See specific types Electro - ocu1ogram (EOG ), 167 , 167 (EEG , 171 ), 92 , Electroencephalogram - 315 228 , 313 (EMG , 167 ), 92 Electromyelogram EMG , 92, 167 , processing Energy requirements capabilities and , 51- 52 EOG , 167 , 77 Epilepsy - 301 - 315 , 81- 82 , 300 , 313 , GABA neurons memory Episodic . 34. 37. 70. 123- 124. 144364 , 372 145. 147. 150. 332 , 357 Gated dipole model. 343- 344 Equipotentiality approach ERPs , 61 , 66, 84 , 93 . 371 Gating function Errors . See also , 66, 396 , 404 General Problem Solver specific types (GPS ) program . 211 Ethanol . See Alcoholdependence , chronic Generate uses exposure task . 65- 67 - 35 Glutamate . 34 -related Event ERPs ( ), 61, 66, 84, 93 GPS . 211 potentials program , 235 Gradient -descent Evolutionary theory . 12 methodology Executive functions , 25 , 65- 68, 70 , 209 , 365 - 388 366 HO, 383 , 391 Executivenetwork 6568 70 . , Hebbian rule , 302 Hebbian modification Explidt memory, 77, 81- 82 , 322 synaptic 's Extinction Hebb Law 8 113 , , Learning . See Learned explaining , 259 Helplessness helplessness , 259 sensory model depression computational - 264 simulation of, 266- 267 , 257 , 263 Hemineglect

421

Index

's disease - 388 , 391 ( HD ), 383 Huntington HVA . 141


, 37 Hypothalamus

Hiddenlayer , 225 Hiddenunits ,5 -layerweights Hidden , 219 resolution , High electroencephalography 228 fomaation , 80 Hippocampal modelof freeandcuedrecall Hippocampal modulation , 311 cholinergic - 306 and models connectionist , 305 localnetworks , 304 305 - 320 in, 311 function memory - 312 cuedrecallin spikingmodel , 311 - 320 on freerecall , 319 diazepam - 315 , 313 memory episodic freerecallin continuous , 317 firing model 318 - 317 assodate learning , 315 paired on free recall 319 - 320 , scopolamine and 306 310 , subregions between , 306 connecting of, 306 - 310 dynamics of, 310 - 322 modification , 320 synaptic - 324 useof, 299 , 323 subregions Hippocampal between , 306 connecting of, 306 - 310 dynamics Cortico , 80. Setalso system Hippocampal andmemory system hippocampal , 46- 47, 80 Hippo campus Uicacid( HVA Homovani ), 147 Tank model 153 , Hopfield Human , 33, 61 cognition model Human memory - 321 ' s disease Alzheimer and . 320 300 models 299 , computational - 324 , 323 findings modelof freeandcuedrecall , hippocampal - 311 304 modulation , 311 cholinergic between , 306 subregions connecting - 306 models connectionist , 305 - 310 of subregions , 306 dynamics - 305 localnetworks , 304 - 320 function in, 311 memory - 322 of modification , 310 , 320 synaptic - 324 useof model , 299 , 323 - 323 model neocortical function , 321 neural networks and . 299 , 301 304

Brain imaging types ; specific Imaging. See Implidt memory, 77, 81 82 . SetCognition Information-processing nodes 363 , Input Interadive architecture , modular architecture versus , 19- 21 Interadivity , 397 Interlevel resonantfeedback . 359, 361 Intermediatetransient memory ( ITM) bilateral lesionsand , 90 term , 99 memory and long , 92 partial lesionsand short-term memory and , 101 , 104, 107 temporal order network and and 83 79, , , 85, 94, 97, 113 working memory Internal control, loss of, 147 Internal representations , 10, 238- 240 Internal structure of neural networks, 22- 23 Invarianceprindple , 106 ITM . SetIntermediatetransient memory Knowledge, 22, 225, 273, 381 382, 407 411 Kohonen network ,9 ' Korsakoff s syndrome , 91- 92 Languagedelicts , 193 195, 205, 382 Lateral inhibition, 359- 360 LC system , 146, 149 and depressioncomputa Learnedhelplessness tional model affed-driven helplesmess and , 150 , 153- 160 building, in steps " American PsychologicalSociety model in " prindple , 153 154 ongoing work, 156- 160 Washington Evolutionary SystemsSociety ' 95 model, 154- 156 ' World Congresson Neural Networks 94 model, 154 dopamineand, 150 GABA neuronsand , 150 loss of internal control, 147 mastery versuscontrol, 145- 147 overview of, 141- 142 , 145- 150 Sampsonand , 142- 145 types of learnedhelplesmess 147 Welss and, 150

422

Index

Learning associative - 360 , 359 - 343 -based . 342 contingency correlational , 81 cortico - hippocarnpal andmemory , system 77 , 80 deficitin, 87 's law 8 113 Hebb , , -tennmemory and , 80 long - 340 motor , 338 networks , 9- 10 by neural nontrivial , 22 - 246 and , 245 numerosity - 317 associate , 315 paired rate , 12 search space during , 22 semantic networks and , 10 - 106 , 105 sequence Towerof Hanoitestand - 220 , 219 states and , 182 wakingandsleeping Lemmas , 272 Lesion studies , new - 289 for actions , 287 - 280 for nonunique concrete entities , 275 - 285 for unique concrete entities , 283 - 259 - 267 Lettercancellation , 257 , 265 Lex .icalmediation units , 272 Lexical retrieval for actions , 286- 289 , 286- 287 background , 289 findings functional (PEl ) studies , 289 imaging - 289 lesions studies , new , 287 of, 290 - 291 findings - 273 neural networks , 290 supporting , 271 291 - 282 for nonunique concrete entities , 273 - 275 , 273 background - 282 , 281 findings functional - 281 (PET ) studies , 280 imaging - 280 lesionstudies , new , 275 , 273 terminology - 286 for unique concrete entities , 282 282 283 , background - 286 , 285 findings functional (PEl ) studies , 285 imaging - 285 lesion studies , new , 283 Umbicnetworks -based - 343 contingency learning , 342 's disease 346in Parkinson and , depression 347 - 346 disorders and , 343 neuropsychiatric , 346 Strooptestand

- 103 102 -tennmemory LTM. See Long

strudure, 332, 342 Trail Making test and , 346 Une bisection , 257- 259, 265- 267 Lobotomy, 48 Loa of processingimpairments , 395 Locuscenaleus (lO system , 146, 149 Long-term memory (l TM ) amnesiaand, 86- 87 bilateral lesionsand, 90 in cortica- hippocampalsystem and memory, 78- 80, 112- 113 in functional model of hippocampo -cortical relations , 98- 103 intermediatetransient memory and, 99 , 80 learning and in mathematicalmodel of hippocampo cortical relations , 104 partial lesionsand, 92 transient , 78, 97, 111, 113 , 78- 79, 83, 111 working memory and term Long potentiation (LTP), 79, 83, 97, 99,

M-P neuron ,8 M-Q desaiptor , 214 resonance (MRI), 93- 96, Magnetic imaging 277 functional , 61, 66, 92, 228 MAP-3 technique , 277 - 147 , controlversus , 145 Mastery Mathematical modelof hippocampo -cortical relations , 104 - 249 Mathematical , 246 reasoning , 233 MDSanalysis , 383 , 387 , 389 Medialandorbitalprefrontal cortex - 366 ( MOPFC ), 365 Medialseptum - 310 , 309 . Setalso Cortico - hippocampal Memory system andmemory ; Human model ; memory Recall active , 48- 51, 84, 96 - 303 associative , 301 declarative , 77 , 81- 82 - 315 81 82 - 301 , , 300 , 313 , 364 , episodic 372 , 77 , 81- 82 explicit frontallobeand , 48- 51 , 77 , 81- 82 implicit intermediate transient bilateral lesions and , 90 -termmemory and , 99 long and , 92 partiallesions

423

Index

. Set also Cortico - hippocampal Memory system andmemory model ; Human ; memory Recall .) (cont -tennmemory short and , 101 and , 79 , 83 , 85 , 94 , 97, workingmemory 113 load , 89, 176 -tenn long amnesia and , 86- 87 bilateral lesions , 90 in cortico - hippocampal andmemory system - 80 - 113 , 78 , 112 in functional modelof hippocampo -cortical relations , 98- 103 intennediate transient and , 99 memory and , 80 learning in mathematical modelof hippocampo cortical relations , 104 and , 92 partiallesions transient , 78 , 97, 111 , 113 and , 78- 79 , 83 , 111 workingmemory , 46- 47 , 49- 50 monkey
neural networks and, 34 nodes , 363 overload, 176 , 96 passive , 77, 81- 82 procedural retentive, 49- 51 semantic , 81, 372 short-tenn

intermediate transient and , 79 , 83 , memory 85, 94, 97 , 113 -termmemory - 79 and , 78 , 83, 111 long -termmemory short and , 83 - 246 networks , 245 Memoryandmatching Modulararchitecture interactive architecture , versus , 19- 21 of, 12 Momentum , parameter - 366 MOPFC , 365 - 335 Motor control , 334 Motor learning 338 - 340 , - 338 Motor maps , 337 Motor networks - 335 motorcontrol , 334 - 340 motorlearning 338 , - 338 motormaps , 337 's disease Parkinson and , 333 - 337 mechanisms , 335 pharmacologic - 333 structure , 332 - 143 Motoncdepression , 142 Move , 214 ) desaiptor -Q -quality( M Movement , 37 MRI, 93- 96 , 277 functional , 61 , 66, 92 Multidimensional ( MDS ) analysis , 383 , scaling 387 , 389 - 275 Naming , 273 - 407 , 396- 398 , 400 impainnents nBM , 370 NE - 147 , 143 , 146 , 150 , 152 , 170 , 345 , 66 Negativity Neglect anatomical - 261 network . 260 - 259 assessing , 257 networks and ,8 backpropagation - 261 explaining , 259 Neocortex , 37 - 323 Neocortical function model , 321 Nervous of primates , 33 34 system - 18 NE Ttalkreading model , 17 - 125 Neural , 124 , 234 , 251 circuitry Neural network . See also modeling specific types - 303 associative memories and , 301 clinical research and , 25- 26 - 21 debates in, 15 modular versus interactive architecture , 19 21 - 17 nature of, 15 andrules , need symbolic representations - 17 for, 15 in, 25 development

active , 48- 51 modelof, 106 algebraic amnesia and , 86- 89 in corti co - hippocampal and system - 113 , 78 , SO , 107 , 112 memory deficits in , 89 in functional modelof hippocampo -cortical relations , 98- 103 intennediate transient , 101 memory in mathematical modelof hippocampo cortical relations , 104 and , 91- 92 partiallesions ordernetwork and , 107 temporal and , 83 workingmemory " - 301 ; 300 snapshot states and , 182 wakingandsleeping working automatic , 79 , 83- 86 , 94 brainimaging and , 92- 96 - 80 cortico - hippocampal and , 78 , system - 86 83 , 96- 98 human versus animal , 96- 98

424

Index

historical of, 3- 4 perspective introduction of, to mainstream , 3 limitations of, 21- 25 , 23 biological plausibility interferences , 23- 25 catastrophic internal structure , 22 23 of neuropsychological data , complex , 25- 26 - 17 in, 10 , 15 representations -organizing - 304 self and , 9, 301 systems , - 307 306 337 , - 382 semantic and , 381 knowledge -processing models versus symbol , 5- 15 - 15 resonance , 4, 12 adaptive , 20 , 26 theory attractiveness of neural networkapproach , 6- 7 networks , 4, 7- 12 , 26 backpropagation contrast between traditional andneural networkapproach and , 5- 6 and 15 17 , representations termof, 3 states , 179 wakingandsleeping Neural networks animal and , 19 systems of, 21- 22 applications artiAcial , 8- 10 , 334 attention and , 62- 71 executive , 65- 68, 70 - 71 interaction of, 69 - 70 6265 , 69 orienting , 68 70 , vigilance basal and , 70 , 331 ganglia of, 23 biological plausibility of, 9 components described , 8- 9 - 304 human modeland , 299 , 301 memory internal structure of, 22- 23 by, 9- 10 learning and , 34 memory - 304 overview , 301 - 361 of, 359 principles sizeof, 18 , 21- 22 -lesioned Towerof Hanoitestin frontaiiobe - 228 , 218 patients - 224 of, 221 degradation - 228 evidence from , 224 testing , 221 - 220 training , 218 9, 12 , 17 training , - 19 view of, 18 - 361 Neuromodulation , 359 - 52. See Neurons , 34- 35, 42- 43, 51 also specific types

- 346 disorden . Setalso , 343 Neuropsychiatric sptcific types data , 86- 92 Neuropsychological rehearsal allowed , 87 89 rehearsal , 89- 92 prevented tests Neuropsychological ActionRecognition andNaming Test , 287 - 284 Boston Famous Faces Test , 283 -Petenon Brown task , 319 Buschke test , 370 test , 358 , 373 categorical fluency continuous test(setContinuous performance test ) performance Dementia Scale Rating(setDementia ) Rating , 89, 300 , 319 digit span -Oldfieldquestionnaire Geschwind , 275 IowaFamous Faces Test , 283 lettercancellation (setLettercancellation ) linebisection (setLinebisection ) Test ReyAuditoryVerbalLearning , 372 ) Strooptest(setStrooptest Towerof Hanoi(setTowerof Hanoi ) - 227 Towerof London , 210 , 212 , 225 Towerof Toronto , 212 Trail Makingtest(setTrail Makingtest ) verbalfluency (setVerbalfluency ) Wechsler , 300 , 370 MemoryScale Paired Associate Subtest , 300 , Learning - 313 - 317 311 , 315 Wisconsin CardSortingTest(setWisconsin CardSortingTest ) Neurotransmit ten, 34- 35, 124 . Setalso sptcific types NMDA receptor , 45- 46 , 123 movement ( NREM eye ) sleep , 166 Nonrapid 167 - 172 , 170 , 176 , 180 , 182 Nontriviallearning , 22 concrete entities , lexicalretrieval Nonunique for - 275 , 273 background defined , 273 - 282 , 281 findings functional - 281 (PET ) studies , 280 imaging - 280 lesions studies , new , 275 - 147 ( NE , 146 ), 143 , 150 , 152 , Norepinephrine 170 , 345 - 167 NREMsleep - 172 , 166 , 170 , 176 , 180 , 182 Nucleus basalis of Meynert(nBM), 370 - 238 Number processing , 235 Numerical 238 , quantities - 240 - 246 , 241 Numerosity

425

Index

disorder Obsessive - compulsive (OCO ), 125 126 , 128 , 132 , 136 - 126 OCD, 125 , 132 , 136 , 128 of primates , 34- 35 Olfactorysystem 359 360 , processing Opponent - 366 &ontalcortexmodel Orbitalpre , 365 65 70 , 62- , 69 Orientingnetwork -layerweights , 219 Output
Pairedassociate learning , 315- 317 Paralleldistributed processingmodels 's disease Alzheimer and , 396- 413 architecture , 398- 400 , 407- 411 categoricalknowledge preserved 412 413 future directions , , 411- 412 implications for neuropsychology lexical impairments , 398 400 semanticimpairments , 396- 398, 400- 407 visual impairments , 398- 400 approachof, 395 396 attentional disordersand, 260 , 33, 35 cognitive networks in primatesand term of, 3, 10- 11 Parietallobe, 62- 63 ' with depression(POD), Parkinsons disease 346 ' Parkinsons disease(PO) basalganglia dysfunction and, 333 depressionin, 346- 347 motor control and, 335 motor networks and , 333 and , 335- 337 pharmacologicmechanisms spatial attention and, 341 Partial lesions , 91- 92 Passive memory, 96 Pathfinderanalysis , 387, 389- 391 ' s disease PD. Set Parkinson ' with depression POD (Parkinsons disease ), 346 POD (pervasivedevelopmentdisorder), 145 ron , 8 Percept Perseveration , 150, 152 Pervasivedevelopmentdisorder (POD), 145 PET , 61, 63, 66, 93- 96, 146, 171- 172, 228, 273, 346. SetalsoFunctionalimaging (PET ) studies . SetPrefrontalcortex PFC PGO waves , 170- 171, 176- 177, 180 mechanisms , 335- 337. Setalso Pharmacologic sptcificdrugs Pontine cells , 170 - geniculo - occipital (PGO) waves Ponto , 170171, 176- 177, 180

emission Positron (PET ) , 61, 63, tomography - 172 . 66, 93- 96, 146 , 228 , 273 , 346 , 171 also Functional See (PET ) studies imaging -tetanic Post ( PTP ), 102 potentiation Acalculia and &ontalcortex(PFC Pre ). Setalso models ontaldeficits pre& - 366 dorsal , 365 dorsalateral , 357 - 366 orbital , 365 - 196 - 192 and , 191 , 195 schizophrenia Wisconsin CardSortingTestmodeland to, 20 damage Pre &ontallesions , 248 Pre & , 48 ontallobotomy Procedural , , 81- 82 memory77 , energy Processing capabilities requirements and , 51- 52 of context in , 192 , disturbances Processing 195 rule Pseudonormalization , 106 PTP , 102 cells , 34- 35, 51 Pyramidal - 343 networks associative Random , 342 - 168 , 166 , ) sleep (REM eyemovement Rapid - 177 170 - 173 , 176 Reasoning , 364 modelof freeand . See Riso Recall Hippocampal cuedrecall ; Memory of declarative , 77 memory defidt in, 87 free , 91- 92 - organization self and , 304 zero , 300 - 275 , 360 , 42, 273 Recognition Rehearsal allowed , 87- 89 , 89- 92 prevented - 177 - 168 - 173 REMsleep , 176 , 170 , 166 , , 238 240 , 359 , 10 , 15 17 Representations 397 Resonant , 360 codingandrecognition Retentive , 49- 51 memory - 78 amnesia , 77 , 82 Retrograde Rule -coding clusters , 363 - 321 modification , 320 Runaway synaptic
Schaffercollaterals , 309- 310 Schizophrenia , 191 biological mechanisms connectionistmodels , 191- 192, 206 in processingof context and disturbances . 191- 195

426

Index

- 198 and , 195 dopamine - 196 effects of, 195 neuromodulatory effects of 197 198 , physiological - 196 cortexand , 195 prefrontal frontallobedysfunction and , 196 - 195 deficits and , 205 , 193 language - 192 - 196 cortexand , 191 , 195 prefrontal - 206 semantic networks and , 203 of simulations context - dependent lexical , disambiguation - 206 203 - 203 test continuous , 193 , 201 performance of, 192 observations effects of dopamine , 197 physiological 198 - 200 effect , 193 , 198 Stroop Wisconsin CardSortingTestand , 196 on freerecall - 320 , 319 Scopolamine - 304 -organization Self , 306 network , 9, 301 309 , 337 -recurrent Self , 131 weights elaboration Semantic , 248 - 382 Semantic , 381 knowledge Semantic , 81, 372 memory networkmodel Semantic - 391 Alzheimer 's disease and , 382 deterioration of, 389 - 391 - 382 view of, 381 functional of, 383 388 paststudies and , 383 , 385 cognitive maps and 381 382 , knowledge networks Semantic and , 10 learning - 206 and , 203 schizophrenia extinction , 259 Sensory - 106 learning , 105 Sequence , 179 properties Sequencing in CAl , rapid self , Sequential organization - 309 308 - 124 function , 123 Serotonergic 5 34 52 Serotonin , 170 , 345 ) , , , 146 ( HT facilitation -duration Short (STP ), 99 synaptic -termforgettingtask Short , 91 -termmemory Short ) (STM - 51 active , 48 modelof, 106 algebraic and amnesia , 86- 89 andmemory in cortico - hippocampal , system - 113 78 , 80, 107 , 112 deficits in, 89 in functional modelof hippocampo -cortical relations , 98- 103 transient and intermediate , 101 memory

in mathematical modelof hippocampo corticalrelations , 104 and , 91- 92 partiallesions network order and , 107 temporal and , 83 workingmemory index . 390 Similarity Single tomog photonemission computed 225 226 228 , , ) (SPEC1 raphy states . See Wakingandsleeping Sleeping states " " - 301 , 300 Snapshotmemory eventdetection , 36 Somatosensory of primates , 36- 37, system Somatosensory 47- 48 attention , 341 Spatial - 226 SPECT , 225 , 228 - 111 . 110 timingnetwork Spectral -brainpatients , 237 Split andmodulation State - dependent , sequencing
177- 179 STM. See Short-term memory STP 99 , Stroop effect, 193, 198 200 Stroop test as attentional task , 67- 68 backpropagationnetworks and, 8 disturbancesin processingof context and , 192- 193 with , 198 empirical phenomenaassociated limbic networks and, 346 Subordinatelevel, 273 Subregionsof hippocampus , 306 connecting between of 306 310 , dynamics , 63- 64 Superior colliculus . 65 Supervisory system , neural network Symbol-processingmodels as alternative to modeling theory, 4, 12- 15, 20, 26 adaptive resonance attractivenessof neural network approach and, 6- 7 backpropagationnetworks, 4, 7 12, 26 contrast between traditional and neural network approach , 5- 6 and, 15- 17 representations and rules, need for, Symbolic representations 15- 17 , 310, 320- 322 Synapticmodification Synaptic triads, 361 Temporal order in functional model of hippocampo -cortical relations , 99

427

Index

TrailMaking test , .346 " " model - 2.38 -code .35 ,2 Triple

order(cont .) Temporal intermediate transient and , 104 , 107 memory - 108 networkmodels , 104 - 80 of serial , 79 sequences asspatially distributed , amplitude gradient 106 - 107 - 108 ordernetwork(TOM), 104 Temporal of es , Temporal sequencing analytic process 43- 45 - 111 Timing , 109 TOH. See Towerof Hanoi - 108 TOM, 104 Towerof Hanoi(TOH also Neural ). See networks - 12 networks and , 4- 5, 11 backpropagation in classic form , 11 -lesioned - 228 in frontaliobe , 209 patients of, 209 completion - 214 moves , 212 encoding and 219 220 , learning - 228 neural networks and , 218 dataand , clinical , 209 neuropsychological 210 - 217 , 214 selecting trainingset - 212 for, 210 solving , strategies and - 220 , 219 learning

Unilateral lesions , 87 (UD), 346 majordepression Unipolar concrete entities retrieval for , lexical Unique - 283 , 282 background defined , 273 - 286 , 285 findings functional (PET ) studies , 285 imaging - 285 lesions studies , new , 283 Unlearning , 176 -learning net ,9 Unsupervised - 144 , 142 Vegetative depression Ventralpathway 41 42 , 'Verb " , 289 generateparadigm Verbalfluency model - 369 , 366 - 373 Alzheimer 's disease and , 369 68 70 network , Vigilance Visualeventdetection , 36 Visualsystem of primates , 36- 37, 40- 47 , 84 Visuospatial sketchpad , 248 Visuospatial workingmemory states Wakingandsleeping - 166 of, 165 categories

- 168 and , 165 cognition - 177 model , 172 computational brainstem andthalamic models , 173 - 177 distributed andcortical model , 173 overview 172 , results , 182 and , 172 depression considerations - 172 , 166 experimental - 168 , 167 cognitive properties features - 167 , 166 general - 172 data , 168 neurophysiological illness , 172 neuropsychiatric and , 182 learning and , 182 memory neural networkmodeling , 179 simulation results - 181 , 179 state - dependent andmodulation , sequencing - 179 177 ' Washington Evolutionary Systems Society95 model 154 156 , WCSTmodel . See Wisconsin CardSorting Testmodel , 8- 10 , 12 , 131 , 219 , 342 Weights -Pairs -Sumtests WilcoxonMatched Rank , 130 Wisconsin CardSortingTest( WCST ) acalculia andprefrontal deficits models and , 249 - 251 - 358 frontallobefunction modeland . 357 , 48 lobotomyand prefrontal and . 196 schizophrenia Wisconsin CardSorting Test( WCST ) model - 362 resonance . 361 theoryand adaptive of, 20 components - 364 frontallobefunction modeland . 361 cortex and prefrontal damage , 20 verbalfluency modeland , 366 WM. See Workingmemory Word forms , 273 Word retrieval . See Lexical retrieval ( WM ) Workingmemory automatic , 79 , 83- 86, 94 brainimaging and , 92- 96 - 80 cortico - hippocarnpal and , 78 , 83system 86 human versus animal , 96- 98 intermediate transient and , 79 , 83, memory 85, 94 , 97 , 113 -termmemory - 79 and , 78 , 83 , 111 long -termmemory short and , 83 '94 WorldCongress on NeuralNetworks model , 154
Zero recall ,

428

Index

También podría gustarte