Conference on Cognitive Computational Neuroscience

Understanding how the brain encodes information requires knowledge of cortical activity structure, including single-neuron variability and shared covariability between neurons. However, studies of visual-cortical representations of natural scenes have mainly focused on single neurons, therefore the role of covariability in natural visual processing remains poorly understood. We hypothesize that the structure of trial-to-trial covariability in primary visual cortex (V1) approximates an optimal representation of natural visual inputs, thus relating V1 population activity structure and scene statistics. We model V1 activity as representing probabilistic inferences via neural sampling in a well-known generative model of image statistics. We show that when multiple neurons share a global feature, such as image contrast, large stimuli reduce noise correlation. However, the correlation increases between neurons that do not share global features. Further, given the receptive field properties of two neurons, natural image statistics determine which of these two models is a better predictor of their noise correlations. Our results thus capture diverse patterns of context-dependent modulation of covariability and offer new experimentally testable predictions.