Biological constrained learning of parameters in a recurrent neural network-based model of the primary visual cortex

Abstract

Neurons in primary visual cortex (VI) optimally respond to stimuli with their preferred orientation. The response of neurons in VI is suppressed by iso-oriented neurons located in their surround. It is very important to understand the circuitry of center-surround interactions. Previous studies in this field followed the approach of postulating models inspired by neuroscience data. While previous models are only postulated, we adopted a strictly data-driven approach and trained a biologically constrained recurrent network model by using supervised learning methods. We have trained a recurrent neural network model constrained by selected biological and anatomical facts. The obtained model describes the near and far surround behavior and the synaptic weights obtained by training are biologically plausible.