Item

Abstract

Contribution of joint statistics of neuron populations to stimulus encoding can distinguish theories of neural computation. Specifically, probabilistic inference in a hierarchical model of perception predicts the emergence of content-specific modulations in the fine structure of spike count correlations. By recording spiking activity from the V1 of behaving macaques viewing naturalistic and synthetic stimuli we demonstrate that compositional objects elicit correlational structures that are more specific to the identity of the stimulus than stimuli without structured content. Further, we demonstrate that decoding schemes exploiting stimulus-specific pairwise response statistics outperform those relying on marginal statistics, thus showing that joint statistics carry information about the stimulus independently from marginal statistics. To rule out possible simpler explanations of the observed patterns in the correlation structure, we introduce an array of controls. We develop Contrastive Rate Matching to control for firing rate-related changes in correlation magnitudes. Further, we analyze phenomenological models of noise correlations, the raster marginal model and a family of models featuring collective additive and/or multiplicative noise sources. Our results show that stimulus-dependence of noise correlations at the level of V1 reflect high-order structure in the stimulus, is independent of changes in firing rates and cannot be explained by phenomenological accounts.