Computational neuroengineering for artificial retina
A major goal of neuroengineering is to control spiking activity at cellular resolution in a way that replicates neural code. Our context is the artificial retina, a neural interface that aims to treat blindness due to retinal degeneration by interfacing with remaining retinal ganglion cells. Using primate retina recordings from a large-scale multielectrode array as a lab prototype of the device, we study three important problems : (1) How does the retina normally respond to visual stimuli? (2) How do we use healthy retinas to predict responses in a blind retina, in spite of individual variability? (3) How can we replicate this response pattern by electrical stimulation?
First, we build a model for visual encoding that incorporates non-linear spatial integration, a prevalent computational motif in neural circuits. We identify subunits that hierarchically partition the receptive field into localized components. For prosthesis application, we extend the method to find shared nonlinear substrates in cell populations and describe their role in explaining responses to natural stimulus.
Second, for translating these models to restoring visual function in a blind retina, we study the variation in visual encoding across a hundred recordings from different animals. We learn a low dimensional manifold to separate individual variability from shared computations. Using the manifold, we study the covariation of response properties of different cell-types and observe surprising differences in neural encoding between males and females. The low dimensional manifold also helps in efficiently identifying the neural code in a previously unseen retina, and calibrating device function for a blind retina.
Finally, we present an approach to replicate neural responses as accurately as possible when the interface has imperfect selectivity. By temporally multiplexing a collection of non-selective patterns, we show improved performance compared to the algorithms used by existing devices. Under simplifying assumptions, a greedy algorithm for identifying the sequence of stimulation is surprisingly close to optimal. This framework allows us to use experimental data to derive insights on the hardware design.