Models of Visual Processing by the Retina

Abstract

The retina contains neural circuits that carry out computations as complex as object motion sensing, pattern recognition, and position anticipation. Models of some of these circuits have been recently discovered. A remarkable outcome of these efforts is that all such models can be constructed out of a limited set of components such as linear filters, instantaneous nonlinearities, and feedback loops. The present study explores the consequences of assuming that these components can be used to construct models for all retinal circuits. I recorded extracellularly from several retinal ganglion cells while stimulating the photoreceptors with a movie rich in temporal and spatial frequencies. Then I wrote a computer program to fit their responses by searching through large spaces of anatomically reasonable models built from a small set of circuit components. The program considers the input and output of the retinal circuit and learns its behavior without over-fitting, as verified by running the final model against previously unseen data. In other words, the program learns how to imitate the behavior of a live neural circuit and predicts its responses to new stimuli. This technique resulted in new models of retinal circuits that outperform all existing ones when run on complex spatially structured stimuli. The fitted models demonstrate, for example, that for most cells the center--surround structure is achieved in two stages, and that for some cells feedback is more accurately described by two feedback loops rather than one. Moreover, the models are able to make predictions about the behavior of cells buried deep within the retina, and such predictions were verified by independent sharp-electrode recordings. I will present these results, together with a brief collection of ideas and methods for furthering these modeling efforts in the future.