Deciphering the neural code for retinal ganglion cells through statistical inference

Abstract

This work studies how the visual system encodes information in the firing patterns of retinal ganglion cells. We present a visual scene to a retina, obtain in-vitro recordings from a multi-electrode array, and attempt to identify or reconstruct the scene. Our approach uses the well-known linear-nonlinear Poisson model to characterize neural firing behavior and accounts for stochastic variability by fitting parameters using maximum likelihood. To characterize cells, we use white noise analysis followed by numerical optimization to maximize the likelihood of the experimentally observed neural responses. We then validate our method by keeping these fitted parameters constant and using them to estimate the speed and direction of moving edges, and to identify a natural scene out of a set of possible candidates. Limitations of our approach, including reconstruction fidelity and the validity of various assumption are also examined through simulated cell responses.