Está en la página 1de 3

# Dimensionality Reduction of Neural spike train data using factor analysis, including a comparative review of linear Gaussian models

If we can get the low-dimensional features of the neural signals by doing dimensionality reduction, pattern classification tasks could become easier. [Key idea] A generalization of PCA, Factor Analysis (FA) handles distinct noise variability in each neuron, making it more appropriate for neural data. What is PCA? It finds projections in the high-dimensional space, then throwing out the less-informative ones.

PCA finds dimensions with maximum variance of projection. PCA finds dimensions that minimize error of projection. Both lead to a solution based on eigenvector decomposition. Probabilistic PCA (PPCA) [Key Idea]PCA can also be expressed of as maximum likelihood estimation of a probabilistic continuous latent variable model [pattern recognition and machine learning, Bishop, Section

12.2]. Each data point Xn, can be described by a latent variable Zn in the space Z, which is generative. Factor Analysis (FA) [Key idea]FA is an extension of PPCA, in which the covariance structure of the noise is less constrained. In PPCA, the covariance matrix corresponding to input noise must be a multiple of the identity matrix, , is the global noise level. For lim0 , it is vanilla PCA. For FA, the covariance matrix is diagonal. FA retains distinct noise variance in each dimension. Equivalence of FA and Mixture Models Gaussian factor model with k factors is equivalent to some mixture model with k +1 clusters, in the sense that the two models have the same means and covariance. *Shalizi+ factor analysis cant really tell us whether we have k continuous hidden causal variables, or one discrete hidden variable taking k+1 values *Shalizi+ [KEY] Analysis of Neural Data Assumptions: 1. Each neuron is a noisy sensor reflecting temporal evolution of a low-dimensional internal process. 2. Firing rate is sufficient to characterize activity Challenges: 1. Neurons are highly variable: Hence FA is necessary, allowing distinct noise models, especially to normalize for loud vs. quiet neurons 2. Must smooth binned firing rates: Use a two-step process: first smooth, then fit; Authors combine two steps into Gaussian Process Factor Analysis (GPFA) Neural Time Series Analysis - Steps

The goal is to determine trajectory from a single trial. [KEY] Problems with two-step process/traditional approach Kernel smoothing technique is ad-hoc Same kernel for each neuron implies single timescale, probably erroneously PCA has no noise model, making it difficult to isolate noise from data

## No interaction between two phases of process Gaussian Process Factor Analysis

Conclusions and Future Work GPFA offers a powerful tool with several advantages: FA allows explicit noise model for each neuron GPFA extends static model for time-series analysis Use of parametric GP covariance permits extensiveexploratory modeling Future Work: Richer GP neural state evolution Non-stationary kernels Non-linear manifolds and point-process likelihood

Reference Single-trial analysis of neural population activity during motor preparation Dimensionality Reduction of Neural spike train data using factor analysis, including a comparative review of linear Gaussian models (PPT for CS 545, machine learning)