HIERARCHICAL NEURAL COMPUTATION IN THE MAMMALIAN VISUAL SYSTEM

Our visual system can efficiently extract behaviorally relevant information from ambiguous and noisy luminance patterns. Although we know much about the anatomy and physiology of the visual system, it remains obscure how the computation performed by individual visual neurons is constructed from the neural circuits. In this thesis, I designed novel statistical modeling approaches to study hierarchical neural computation, using electrophysiological recordings from several stages of the mammalian visual system. In Chapter 2, I describe a two-stage nonlinear model that characterized both synaptic current and spike response of retinal ganglion cells with unprecedented accuracy. I found that excitatory synaptic currents to ganglion cells are well described by excitatory inputs multiplied by divisive suppression, and that spike responses can be explained with the addition of a second stage of spiking nonlinearity and refractoriness. The structure of the model was inspired by known elements of the retinal circuit, and implies that presynaptic inhibition from amacrine cells is an important mechanism underlying ganglion cell computation. In Chapter 3, I describe a hierarchical stimulus-processing model of MT neurons in the context of a naturalistic optic flow stimulus. The model incorporates relevant nonlinear properties of upstream V1 processing and explained MT neuron responses to complex motion stimuli. MT neuron responses are shown to be best predicted from distinct excitatory and suppressive components. The direction-selective suppression can impart selectivity of MT neurons to complex velocity fields, and contribute to improved estimation of the three-dimensional velocity of moving objects. In Chapter 4, I present an extended model of MT neurons that includes both the stimulus-processing component and network activity reflected in local field potentials (LFPs). A significant fraction of the trial-to-trial variability of MT neuron responses is predictable from the LFPs in both passive fixation and a motion discrimination task. Moreover, the choice-related variability of MT neuron responses can be explained by their phase preferences in low-frequency band LFPs. These results suggest an important role of network activity in cortical function. Together, these results demonstrated that it is possible to infer the nature of neural computation from physiological recordings using statistical modeling approaches