| |
Abstract:
We believe that understanding neural processing will
ultimately require observing the response patterns and interaction
of large populations of neurons. To that end, we have developed a
chronic multi-electrode implant for long term simultaneous
recording from multiple neurons. The task of analyzing firing
patterns, across time and between different units, arising from
this electrode array is a critical and technically challenging
task. We discuss a greedy, incremental algorithm from the
"Helmholtz machine" family that we are using for automated
discovery of ensemble neuronal events. We construct a generative
model that attempts to maximize a bound on the likelihood of
observed firing patterns. The model is constructed incrementally;
each added unit in the model attempts to greedily maximize the
likelihood. While not globally optimal, this strategy is
computationally tractable. We show encouraging benchmark data on
artificially constructed spike trains and promising early results
on some real data.
|
