A Nonlinear Network Model with Application to Modeling the Retinal Responses

Understanding many of the complex visual functions in the vertebrate retina requires characterizing various types of interactions between its excitatory and inhibitory neural pathways. Although quite successful in predicting retinal responses to simple synthetic visual stimuli, existing computational models are not sufficient to characterize retinal responses to more complex visual stimuli whose processing involves various types of excitatory and inhibitory interactions. In this study, we develop a computational model in the convolutional neural network framework in order to incorporate both excitatory and inhibitory computations generating retinal responses. The successfully predicts the spiking rate of retinal ganglion cells. The recovered model parameters represent the linear and nonlinear processing across different layers in the retina. This novel model of retinal responses provides a powerful tool for understanding how complex computations at the integration of excitatory and inhibitory pathways lead to the retina's complex visual functions, which can in turn advance existing retinal prosthetic devices to be able to implement more naturalistic processing of real-world visual scenes.