Leslie B. Cohen University of Texas at Austin Invited debate in session, "Can Babies Do Math?" Presented at 2002 Meeting of the International Conference on Infant Studies. Preparation of this paper and some of the research presented in it was supported by NIH grant HD-23397 to the author from the National Institute of Child Health and Human Development. 2 Can Infants Really Add and Subtract? Can infants really add and subtract? No. I don't think so. At least I haven't seen any convincing evidence of it. However, it can be instructive to examine this claim in some detail. We can do so at three different levels of magnification. To use a camera analogy, we can attach a wide-angle lens and try to see how the claim fits into the panorama of views of infant cognition presented over the last several decades. When I consider these views and how they have changed over the years, it reminds me of a series of mini Kuhnian revolutions or paradigm shifts (Kuhn, 1962). In the late 1960's and early 1970's we had the Chomskian revolution, of course, with its emphasis on both the innateness and modularity of language. The task was to derive a complete description of language based upon a set of predetermined universal rules or principles. About a decade later in the late 1970's and early 1980's we find this same approach extrapolated to explanations of infant cognition. Creative investigators such as Liz Spelke, Alan Leslie, etc. emphasized the early competence of infants, and both the innateness and modularity of infant knowledge. Once again processes regarding development change were de- emphasized. The task was, and still is, for some investigators, to discover the set of innate, predetermined rules infants use to understand the world around them. Karen Wynn's claim that 5-month-old infants have the native ability to add and subtract certainly fits within this theoretical perspective (Wynn, 1995). 3 But about 10 years later, in the mid to late 1980's and early 1990's, another mini theoretical revolution began to appear. From a developmental point of view, these changes could first be seen most clearly with respect to infant language acquisition. In the 1990's and even up to the present time this view has gradually expanded to include infant cognition as well. Kuhn argued that theoretical revolutions often originate from outside the field. That certainly seems to be the case here. This new approach is based upon an amalgamation of cognitive neuroscience, dynamical systems, and computational modeling. The PDP books in 1986, (McClelland & Rumelhart, 1986; Rumelhart & McClelland, 1986), the "Rethinking Innateness" books by Elman et al. in 1996, the new book by Tom Shultz (in press) entitled "Computational Developmental Psychology" are all examples of this exciting new trend. So are the many recent connectionist models on a variety of topics in infant cognition. These models, which range from infant categorization, to object permanence, to word learning, to causal perception, all provide evidence of the viability of this new approach. And one of the most appealing aspects of the approach to me is that it provides a renewed emphasis on mechanisms of change, learning, adaptation, and development. So we could debate whether infants can add and subtract at this global theoretical level. If we did, I would ask proponents of the innate math module, where the module comes from and how it develops? I would ask whether there is one module for mathematical ability or separate modules for addition and subtraction. What about multiplication and division or solving differential 4 equations? And what about domain general mechanisms that seem to apply to this ability but don't fit into a modular framework? But a debate at this level would probably degenerate into an argument over nature versus nurture. We've had these debates before, and there may not be much point in repeating them here. Instead, we could switch to a normal, 50 mm lens and focus our debate on specific groups of studies that one way or another relate to infant addition and subtraction. Karen's talk today was at this level. If we did, I would be the first to admit that several other examples in the literature, like Karen's, purport to demonstrate that infants can add and subtract. The authors' explanations of the underlying mechanisms vary somewhat, but most of these studies do tend to replicate, at least partially, Karen's original results. (But see Wakeley, A., Rivera, S. and Langer, J. (2000) for a failure to replicate). I find it interesting that the studies most closely related tend to have tested older infants, usually 8 or 9 months of age rather than the 5-month-olds Karen tested. I personally believe there is a big difference in how a 5-month-old and a 9- or 10-month-old will perform on these tasks. And I would ask the authors of these other studies why they decided to test older infants, and whether they first tested 5-month-old infants, but failed. But continuing at this 50 mm lens level, there are other relevant studies beyond those specifically dealing with infant addition and subtraction. For example, evidence is mounting that infants may not actually be enumerating objects. Instead they may be responding just to the overall quantity or amount, or in the case of temporally sequential events, to rhythm. A recent Psychological 5 Bulletin article by Mix, Huttenlocher, and Levine (in press) reaches this conclusion after reviewing much of the literature on infant cross-modal matching, subitizing, and addition and subtraction. But if infants are just responding to overall amount, they can't actually be counting items, a prerequisite for adding and subtracting. This overall quantity view suggests, although it doesn't necessarily imply, that infants in the addition-subtraction task may be responding at an ordinal level, to differences such as greater than or less than, rather than to a specific number of items. While we remain at the 50 mm lens level, another relevant body of evidence seems to have been ignored by those investigating infant arithmetic. Since the 1970s infant researchers have found that under certain conditions infants will respond more to a familiar display than to a novel display. We know for example that younger infants are more prone to show a familiarity preference than older infants, and that complex events, or ones that are difficult to process, are more likely to lead to a familiarity preference than simple events. Hunter and Ames (1988) summarized these relationships in the following figures: Insert Figures 1 and 2 Here The dip in their 3-dimensional model represents a familiarity preference. As one can see from the model, the dip is more pronounced the younger the infant and the more difficult (or complex) the task. Let us analyze Wynn's addition and subtraction task from this point of view. Karen's addition and subtraction study tested relatively young infants in a 6 procedure that certainly was complex, with a Venetian blind going up and down, objects moving in and out, a screen raising and lowering, a variety of noises, a hand appearing and disappearing, etc. If one considers Wynns study from this perspective it is at least possible that her infants were showing a familiarity preference rather than surprise to the incorrect number of items in her addition or subtraction task. In order to examine the plausibility of this familiarity preference option we have to switch to a higher level of resolution, or to continue our analogy, to a what would be a telephoto lens. Wynn (1992) did not use habituation, but she did use a violation of expectation task. One can summarize her procedure in a couple of different ways. In her addition task infants first see one object placed on a stage, then a hand goes behind a screen to add a second object, then the screen comes down to reveal either one or two objects. Infants look longer at one than at two presumably because it is the wrong number. But at our current level of resolution, one also can describe the procedure slightly differently. Infants first see one object, the hand does something; then in the test the infant sees one object. But there are six test trials. So on trial two, the infant sees one object, then in the test, 2 objects; on trial 3, one object, then one object; on trial 4, one object, then 2 objects; on Trial 5, one object, then one object; and on trial 6, one object then two objects. Over the course of the experiment infants see one object 9 times and two objects 3 times. If infants have a familiarity bias, they should look longer at one object than two objects, not 7 because one object is incorrect, but because it is more familiar. A similar argument can be made about infants looking longer at two objects than one object on a subtraction problem. They see 2 objects (the incorrect number) 9 times and 1 object (the correct number) 3 times. Katherine Marks and I (Cohen & Marks, 2002) attempted to disentangle these various explanations. We replicated the procedure used by Wynn (1992) with one major exception. Instead of repeatedly testing the infants with one and two items, we tested them with zero, one, two, and three items. We felt that change was necessary in order to differentiate an addition-subtraction explanation, from an overall quantity explanation, from a familiarity preference explanation. The logic behind this manipulation is shown in Figure 3. Insert Figure 3 about here As you can see from this figure, each explanation leads to a different set of predictions. All three explanations lead to the same predictions regarding outcomes 1 and 2, but they lead to different predictions regarding outcomes 0 and 3. We ran three different experiments. Each experiment had two blocks of trials and each block had 0, 1, 2, and 3 as outcomes presented in a random order. Our first experiment was an attempt to replicate Wynn, (1992). The results are shown in the top portion of Figure 4. Insert Figure 4 about here If you look at just the gray bars, which are the 1 and 2 outcomes, our results do replicate hers. In the addition condition infants look longer at 1 item than at 2 8 items, and in the subtraction condition they look longer at 2 items than at 1 item. 1 These results fit any of the three explanations. If you look at the overall pattern of results including outcomes 0 and 3, it looks as though there also is a general tendency to look longer the more items are on the stage. That tendency was not predicted by any of the explanations. The tendency to look longer when there are more items was confirmed by our second study, a control experiment in which the infants just got the test trials. They also looked longer when more items were on the stage, but that result occurred more in the second than the first block of trials. Our third experiment was the most informative. In that experiment we wanted to simulate the infants' exposure to the objects but not have any addition or subtraction manipulation. So we familiarized infants to either one or two items for approximately 2 s before each test. In the simulated addition condition on each trial infants would see one item for 2 s before the screen would come up and then go down to reveal 0, 1, 2, or 3 items. In our simulated subtraction condition infants would see two items for 2 s before being tested with 0, 1, 2, or 3 items. The results are presented in the bottom portion of Figure 4. As you can see, these results look remarkably like those from our first experiment. But remember there was no addition and subtraction manipulation, only familiarization. Neither the directional, nor the addition-subtraction explanation would predict these results. They are only consistent with the familiarity preference explanation. 9 Although our findings don't prove that Wynn's babies weren't adding and subtracting, they indicate to me that her infants could have been just showing a familiarity preference. At the very least they indicate that familiarity preferences must be taken seriously and controlled for in these types of experiments. In the event you want to examine these experiments in more detail, you should know that they will soon appear (or have just appeared) in Developmental Science followed by a set of commentaries, including one by Karen (Wynn, 2002) and a response by myself (Cohen, 2002). In those commentaries the most potentially telling argument against generalizing from our experiments to Wynns is that we used four different types of test trials and she only used two types. Wynn (2002) argued that four outcomes may increase infants attention to perceptual features or teach the infants to expect the unexpected. Another commenter, Carey (2002), also doubted that we would have obtained the same results if we had used two rather than four outcomes. She suggested that perhaps the infants were becoming overloaded by the complexity of the four-choice outcome. Neither commenter explains though, why that overload would lead us to obtain the same results as Wynn in our first experiment. Nevertheless, we found a way to test these counter arguments (Cohen, 2002). We went back to our first experiment and looked just at the first two trials. Having been given only two trials the infants had no opportunity to learn to expect the unexpected, or to have become overloaded. Fortunately we had tested 80 infants in this original experiment, so we had a sufficient amount of data to work 10 with. We now ran two different types of analyses. (Please bear with me; this may get a bit complicated.) In one analysis we tested the notion that infants were looking longer at impossible than possible outcomes. Forty of the infants in their first two trials had received one possible and one impossible outcome. That is, in the addition condition they saw two items (possible) and either 0, 1, or 3 items (impossible). The other 40 infants saw two impossible outcomes in these first two trials. Our analysis examined whether there would be a significant interaction between those infants who saw possible-impossible outcomes and those who saw two impossible outcomes. Would infants look longer at the impossible than the possible outcome, but look equally at the two impossible outcomes? The answer was no! The F value was less than 1, F (1,78) = .077; p >.75. We then went back and did the same analysis a different way. In this analysis we tested whether infants looked longer at familiar than at unfamiliar outcomes. Forty of the infants had received one familiar and one unfamiliar outcome on their first two trials. (One item would be familiar in the addition condition and two items would be familiar in the subtraction condition.) The other 40 infants had received two unfamiliar outcomes. This time the interaction was significant, F (1, 78) = 3.97; p < .05. Infants looked longer at the familiar than unfamiliar outcome, but equally at the two unfamiliar outcomes.
Thus, even when we examined just the first two trials, we found evidence that the infants would look longer at familiar than unfamiliar outcomes, but not that they would look longer at impossible than possible outcomes. 11 So what does this all mean? In order to answer this question we need to switch lenses once again and go back either to our 50 mm lens or our wide-angle lens. If it just means that the original Wynn study had a possible confound and should have been better controlled, there would be little reason to debate. On the other hand if we see the Wynn study as an example from a larger set, a set that includes numerous studies presumably demonstrating the precocity of young infants, then we may have something very significant to debate. I believe our experiments on addition and subtraction, the third experiment in particular, may be telling us something important about these other studies as well. Many of these studies are like Karen's. They use a "violation of expectation" paradigm. They present complicated events with moving, interacting objects to young infants. And more and more they are failing to habituate infants thoroughly prior to testing the infants' violation of expectations. In other words they are setting up situations optimal for producing a familiarity preference. These studies need to be examined (or re-examined) carefully to determine which, if any, may be confounded by this familiarity preference, and future studies of this type need to include controls to rule out such a preference. What else might our third experiment be telling us? Most people who believe in the precocity of young infants also believe that infants have the equivalent of Stage 4 object permanence by 3 or 4 months of age. That is, at a minimum they assume that young infants "know" that when an object is totally occluded it continues to exist. In fact that assumption is often the starting point in 12 the design of studies that involve one object going behind another or being hidden by a screen, as in Karen's studies. Well if these people don't "buy in" to the possibility of a familiarity preference, then Experiment 3 should present a real problem to them. Experiment 3 can be thought of as the simplest possible test of object permanence using the violation of expectation paradigm. Infants just see one or two objects on a stage; a screen is raised to occlude these objects, and then the screen is lowered to reveal either the same number or a different number of objects. From an object permanence point of view, when the number changes, either one or more objects has magically appeared or disappeared. The infants' expectation of object permanence certainly should be violated. As a consequence the infants should look longer at the changed number of objects. But they don't. They look longer at the same number of objects. Either these infants don't really have object permanence, or there is something wrong with the "violation of expectation" paradigm. Finally at the most general level one conclusion upon which we should all be able to agree (Cohen, 2001; 2002) is that we should not ignore earlier theory and research, even studies reported as long as 20 or 30 years ago. What we learned back then can provide valuable insights into topics being investigated today. And it can prevent unnecessary errors. One of the most valuable lessons to be learned is that we should be cautious about attributing sophisticated cognitive processes to young infants when simpler processes will suffice. One such simple process, that really should be considered more seriously, in a wide variety of studies on infant 13 cognition, is a preference for familiarity, particularly when complex events are presented to young infants. 14 References Carey, S. (2002). Evidence for numerical abilities in young infants: A fatal flaw? Developmental Science. Cohen, L.B. (2001, April). Uses and Misuses of Habituation: A Theoretical and Methodological Analysis. Symposium paper presented at the Society for Research in Child Development Meeting, Minneapolis. Cohen, L. B. (2002). Extraordinary claims require extraordinary controls. Developmental Science, 5, 211-212. Cohen, L. B. & Marks, K. S. (2002). How infants process addition and subtraction events. Developmental Science. 5, 186-201. Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D., & Plunkett, K. (1996). Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press. Hunter, M. A. & Ames, E. W. (1988). A multifactor model of infant preferences for novel and familiar stimuli. In C. Rovee-Collier & L. P. Lipsitt (Eds.), Advances in infancy research (Vol 5, pp 69-95). Norwood, NJ: Ablex. Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press. McClelland, J. L., & Rumelhart, D. E. (Eds.). (1986). Explorations in parallel distributed processing. Vol. 2, Psychological and biological models. Cambridge, MA: MIT Press. 15 Mix, K. S., Huttenlocher, J., & Levine, S. C. (2002). Multiple cues for quantification in infancy: Is number one of them? Psychological Bulletin, 128. Rumelhart, D. E. & McClelland J. L. (Eds.) (1986). Parallel distributed processing: Explorations in the microstructure of cognition. Vol. 1, Foundations. Cambridge, MA: MIT Press. Shultz, T. R. (In Press). Computational Developmental Psychology. Wakeley, A., Rivera, S. & Langer, J. (2000). Can young infants add and subtract? Child Development, 71, 1525-1534. Wynn, K. (1992). Addition and subtraction by human infants. Nature, 358, 749-750. Wynn, K. (2002). Do infants have numerical expectations or just perceptual preferences? Developmental Science. Wynn, K. (1995). Origins of numerical knowledge. Mathematical Cognition, 1, 35-60. 16 Footnote 1 . The data in Figure 3 for both Experiments 1 and 3 come from the first block of test trials. Infants received two blocks of trials, but the second block produced no differences. 17 Figure 1: Hunter and Ames (1988) 18 Figure 2: Predicted Looking Times Based upon Three Different, but Possible Explanations Addition Task Familiarity Directional Computational 1 + 1 = 0 Short Long Long 1 + 1 = 1 Long Long Long 1 + 1 = 2 Short Short Short 1 + 1 = 3 Short Short Long Subtraction Task Familiarity Directional Computational 2 - 1 = 0 Short Short Long 2 - 1 = 1 Short Short Short 2 - 1 = 2 Long Long Long 2 - 1 = 3 Short Long Long