Function and circuitry of adaptive inhibition in the retina

The vertebrate retina is comprised of ~100 different cell types that process visual scenes in complex ways. The functions of most of these cells under natural visual processing are unknown, as are the effects of diseases when those cell types are disrupted. Understanding the specific visual functions of retinal circuits has typically proceeded from an identification of an important visual phenomenon such as motion processing, adaptation or prediction, followed by ad hoc experiments that often rely on fortuitous knowledge of pharmacology, physiology, or the random selection of neurons, and necessarily require the use of artificially structured stimuli that have an uncertain relationship to natural scenes. We have developed a highly integrated experimental and computational approach that begins with experiments using natural scenes to create a generalizable and interpretable computational model, and automatically proceeds to testable hypotheses for specific cell types and visual computations under natural scenes. Our approach relies on interpretable neural network models that both capture natural scene responses and a broad range of phenomena of ethological visual processing, and have internal components that carry predictions about the responses and actions of real interneurons. We focus here on a predictive phenomenon that we have previously identified in mammalian and non-mammalian retina known as sensitization, whereby strongly stimulated ganglion cells increase their sensitivity to stimuli that are more likely to occur. In the mouse retina we will test the hypothesis that under natural scenes, sensitization is generated by a diverse set of inhibitory amacrine cells, each of which sensitizes a particular visual feature sensed by ganglion cells. A second goal relates to the critical function of the retina to distinguish between different visual stimuli. Although there has been a great deal of focus on changes in visual sensitivity, the true function of the retina is to support behavior by allowing the higher brain to discriminate between stimuli, from simple determinations of motion direction and object location to complex object recognition. Stimulus discriminability not only depends on the sensitivity of neurons, but also on their stochasticity or noise. We have created neural network models that capture both ganglion cell responses to natural scenes and their stochasticity including noise correlations, allowing us to calculate discriminability for any stimulus condition including natural scenes, and to test experimentally how the actions of interneurons influence discriminability. We study how specific amacrine and ganglion cells influence discriminability. Understanding how visual processing under natural scenes is generated by retinal interneurons is critical to our understanding of retinal diseases involving the degeneration of retinal circuitry. In addition, the computational descriptions of retinal responses under natural scenes will be directly useful in the design of electronic retinal prosthesis systems to best support diverse visual behaviors.