Derivation of natural stimulus feature set using a data-driven model

Abstract

A formal approach for deciphering the information contained within nerve cell ensemble activity patterns is presented. Approximations of each nerve cell's coding scheme is derived by quantizing its neural responses into a small reproduction set, and minimizing an information-based distortion function. During an experiment, the sensory stimulus world presented to the animal is modified to contain a richer set of relevant features, as those features are discovered. A dictionary of equivalence classes is derived, in which classes of stimulus features correspond to classes of spike-pattern code words. We have tested the approach on a simple insect sensory system.