Abstract
The processing capabilities of the visual system certainly depend on “nonlinear” computation at multiple levels in the visual pathway. While successful computational vision models generally recapitulate this, validating and constraining such models using physiological data is confounded by two problems: functional characterizations of visual generally rely on linear “receptive fields” that cannot capture nonlinear effects; and most recordings are from single neurons, implicitly entangling characterizations of their own computations with those taking place in preceding areas. We apply a new nonlinear modeling framework to simultaneously recorded pairs consisting of an LGN neuron and the retinal ganglion cell that provides its main input. Because it is nonlinear, this framework can identify multiple processing elements and their associated nonlinear computations for each neuron. Furthermore, by recording from successive stages of the visual pathway simultaneously, we can distinguish the processing that occurs in the retina from processing that occurs in the LGN, and observe how visual information is successively formatted for the visual cortex. We detect nonlinear processing involving the interplay of excitation and inhibition at both levels. Inhibition in the retina is similarly tuned but delayed from excitation, resulting in highly precise responses in time. Oppositely-tuned inhibition is added at the level of the LGN, whose purpose is less clear, but when combined with inhibition inherited from the retina, likely plays a role in contrast adaptation. Thus, we demonstrate a new method to detect nonlinear processing using easily obtained data at multiple levels of the visual pathway. In doing so, we reveal new functional elements of visual neurons that are generally thought of as mostly linear. This has implications for both our understanding of how information is successively formatted for the visual cortex by its inputs, and suggests more general roles of nonlinear computation in visual processing.