Next: Expectancy driven synchronization Up: No Title Previous: Feature extraction in multiple
Jennifer Linden and Maneesh Sahani
Computation and Neural Systems Program, Caltech
In many neurophysiological experiments, neural responses are recorded during multiple repetitions of a small number of stimulus or behavior conditions. The usual approach to analysis of such data is to consider only the mean response over a defined interval for each trial condition. We take a different approach, asking three questions about the variability of responses both within and between trials. First, how much of the response variability can be attributed to trial condition? Second, given this variability, how much information about trial condition is available in the responses, and (for multiple single-unit data) how does the amount of information depend on the number of cells recorded? Third, what features of the responses carry the information about trial condition?
To address these questions, we examine the discriminability of different trial conditions in multiple spike trains recorded simultaneously from primate posterior parietal cortex. We first reduce the response of the recorded ensemble during each trial to a low-dimensional trial vector. Next, using classifiers of varying complexity, we classify the trial vectors by trial condition and compute the amount of trial vector variability which can be attributed to trial condition. We estimate the performance of the classifier using validation techniques, and obtain from that performance measure a lower bound on the amount of information about trial condition available in the data. Because we are using an explicit classifer (rather than a direct measurement of mutual information), we can then extract and reconstruct the features of the simultaneously recorded spike trains which carry the most information about trial condition. Finally, we repeat our analysis over subsets of the simultaneously recorded cells to assess the relationship between our information estimates and the number of cells included in the analysis. We apply these techniques to sample data collected by the authors in collaboration with John Pezaris and Alexander Grunewald.