Learning Divisive Normalization in Primary Visual Cortex

Abstract

Divisive normalization (DN) has been suggested as a canonical computation implemented throughout the neocortex. In primary visual cortex (V1), DN was found to be crucial to explain nonlinear response properties of neurons when presented with superpositions of simple stimuli such as gratings. Based on such studies, it is currently assumed that neuronal responses to stimuli restricted to the neuron's classical receptive field (RF) are normalized by a non-specific pool of nearby neurons with similar RF locations. However, it is currently unknown how DN operates in V1 when processing natural inputs. Here, we investigated DN in monkey V1 under stimulation with natural images with an end-to-end trainable model that learns the pool of normalizing neurons and the magnitude of their contribution directly from the data. Taking advantage of our model's direct interpretable view of V1 computation, we found that oriented features were normalized preferentially by features with similar orientation preference rather than non-specifically. Our model's accuracy was competitive with state-of-the-art black-box models, suggesting that rectification, DN, and a combination of subunits resulting from DN are sufficient to account for V1 responses to localized stimuli. Thus, our work significantly advances our understanding of V1 function.