A Simulation Study on the Encoding Mechanism of Retinal Ganglion Cell

Understanding how the retina encodes visual information is a key issue for the development of a retinal prosthesis. To study this issue, the neural retina is modeled as a retina module (RM) consisted of an ensemble of spatial-temporal (ST) filters and each ST filter simulates the input-output property of an individual ganglion cell (GC). Two receptive field (RF) models of retinal GC, the difference of Gaussians (DOG) model and the disinhibition (DIS) model, are employed to implement these ST filters respectively. RM performs the encoding operation from an input optical pattern to a group of parallel action potential (AP) trains. To assess the encoding efficiency of RF models, a central visual system module (VM) consisted of a group of artificial neural networks is employed to perform the decoding operation from AP trains to an output perceptual pattern. A matching error is defined as an index to quantify the similarity between the input optical pattern and the output perceptual pattern generated by VM. The simulation results suggest that the matching error declines dramatically when the DOG model is replaced by the DIS model, which implies that the encoding mechanism of the DIS model might be more effective than that of the DOG model.