next up previous
Next: Dynamics of optimal adaptation Up: No Title Previous: Effect of dendritic

A Method For Analyzing Multi-Channel Spike Data

Virginia R. de Sa, R. Christopher deCharms, David T. Blake, and M. Merzenich

Sloan Center for Theoretical Neuroscience
UCSF

We believe that understanding cortical processing will ultimately require observing the interaction of large populations of neurons. To this end, our lab has developed a multi-electrode chronic implant for long term simultaneous recording from multiple cortical locations. The task of analyzing firing patterns across time and between different units arising from this electrode array is a critical but technically challenging task. As the number of neurons sampled increases, the number of possible activity patterns grows dramatically and the data may be too sparse to detect multiple instances of exactly the same pattern. We discuss an algorithm that views the patterns as stochastic and searchs for occurrences of similar patterns within the data rather than for occurrences of identical patterns. The approach is an adaptation of Cascaded Redundancy Reduction, an incremental, ``greedy'' algorithm from the Helmholtz machine family. We map time as a spatial dimension and construct a generative model that attempts to maximize the likelihood of the observed data. The model is constructed incrementally and each added unit to the model attempts to maximally improve the prediction of the data. While not globally optimal, the strategy is computationally tractable, and may serve to find physiologically relevent stochastic patterns within cortical spatio-temporal firing patterns. We show some promising preliminary results with synthetic data.


Tony Zador
Wed Mar 12 22:07:02 PST 1997