Finding likely models that describe population responses

Analyzing the spike response of a neural population will be plagued by a lack of data unless some knowledge of the interrelationships among the discharge patterns of individual neurons places the available data in context. Unfortunately, determining that context a model is precisely what the data analysis must produce. We describe here a new approach to population data analysis that turns the usual data analysis procedure around: Instead of measuring response statistics, such as discharge rates and correlation coefficients, and then finding a single model from them, we find all statistical models consistent with the data. We use an objective information theoretic criterion for determining appropriate models. Our approach generalizes maximum likelihood techniques for estimating model parameters.