Computational Models of Neural Retina

Abstract

SynonymsComputational models of retinal function; Retinal network models DefinitionComputational models of the neural retina simulate the response of the retina to input light. In their most detailed form, the models yield neural output as a spatially varying pattern of spike trains which fully encode the incident dynamic image. In vivo, this retinal output is relayed to higher vision processing centers in the brain such as the lateral geniculate nucleus and the primary visual cortex in humans and primates. Neural retinal models assist not only in the understanding of normal and abnormal retinal function, but are useful for investigating retinal response to artificial stimuli such as electrical stimulation by a vision prostheses. Detailed DescriptionThe vertebrate retina consists of a thin layer of neural tissue which lines the back of the eye. It acts to transduce incoming light signals sensed by rod and cone photoreceptor cells into neural spike patterns output by retinal ganglion cells (RGCs), whose axons together form the optic nerve. The ratio of photoreceptor cells to RGCs is approximately 100:1 (dependent on eccentricity from the fovea), indicating that considerable neural processing occurs in the retina itself prior to the transfer of visual information via the optic nerve to the brain. Each RGC is centered on an approximately circular receptive field of photoreceptors within an associated annular surround. Incident light falling within these regions will have opposite effects on RGC activation, eliciting two main types of response depending on the cell type. On-center RGCs are activated by light falling within their receptive field and inhibited by light falling in their surround, while the reverse holds for off-center RGCs. Both these spatial RGC responses can be mathematically modeled using a simple difference of Gaussians (DoG) (Rodieck 1965).Computational models of the retinal response to light may be broadly classified into integrate and fire, block-structured, and network models. The models vary in complexity from simulating the response of a single retinal output neuron to incident light, through to simulating a spatiotemporal neural response to a 2D time-varying image. Integrate and Fire ModelsThe simplest neural retina models are of the integrate and fire type (Gestri et al. 1980;Reich et al. 1997), shown in Fig. 1 fully generalized to the spatiotemporal domain. For a given input light