Measuring the Information Expressed by Neural Discharge Patterns

Abstract

Various measures have been used to assess how well single neurons represent information. Modeling discharge patterns as stochastic point processes, we determine how well certain measure accomplish this task. We show that the information theoretic measure—capacity—can do a poor job. The mean-squared error measure more accurately describes the fidelity to which sensory signals can be extracted. Calculation of fundamental bounds on mean-squared error show that time-varying signals must have bandwidths orders of magnitude less than the average discharge rate (under a Poisson model) if accurate signal representations are to result. This result indicates that neural ensembles must be considered to understand information encoding by neurons.

Supported by grants from the National Science Foundation and from the National Institute of Mental Health