A Convolutional Neural Network-based Model of Neural Pathways in the Retina.

Abstract

A variety of the visual functions of the vertebrate retina rely on the interactions between excitation and inhibition within specialized retinal circuitry. The nonlinear properties of these interactions, as well as a large number of cell types and pathways involved, pose a challenge for characterizing many of these retinal functions. To address these challenges, we develop a computational model in the convolutional neural network framework, along with an efficient, robust learning procedure that enables incorporating both excitatory and inhibitory interactions between the network layers of the retina through feedforward and lateral connections. The model successfully describes the spiking rate of retinal ganglion cells, for both real and simulated validation data. The recovered biologically-plausible model parameters represent the linear and nonlinear computations across retinal network layers. This novel model of retinal responses and the associated learning algorithm provide a powerful tool for understanding the complex visual computations in the retina and for replicating naturalistic visual functions of the retina for retinal prostheses applications.