Exploring Stimulus-Dependent Cluster Codes in Neocortical Activity Through Independent Mixture Models

Abstract

It is increasingly well-documented that information is generally not encoded through single neurons, but instead through clusters of neural activity which are robust to noise despite extreme variability at the level of their constituent neurons and synapses. Foundational to extrapolating these population codes is identifying the underlying correlational structures which give rise to observable activity patterns. Stimulus-dependent clusters in the retina have been studied extensively, but whether the neocortical populations form similarly-structured clusters is not well understood. Modeling cluster code statistics in the neocortex is a crucial step toward understanding how cortical regions encode, integrate, and process information within behavioral contexts. To this end, independent mixture models generalizing Bernoulli, Poisson, and doubly-stochastic super-Poisson processes were fitted onto data from L2/3 of the primary visual cortex. Despite not capturing the full range of variance in the population, model analysis revealed highly discriminable clusters and a population code which strongly resembles that of the retina, contributing to the growing body of literature surrounding global communication modes neural populations can use to encode information which are based in efficient, unsupervised learning of stimulus-dependent cluster representations.