Toward a unified theory of spatiotemporal processing in the retina

Why do stabilized images fade? How can X cells in the retina respond linearly to a broad range of spatiotemporal stimulation functions in spite of possible nonlinear preprocessing? Why have models of spatial vision that utilize static images and assume linearity of preprocessing led to reasonable results in spite of the two questions above? This chapter introduces the push-pull shunting network, a model of spatiotemporal visual processing that resolves these and other controversial findings. Development of the model is based on an analysis of the spatial and temporal response characteristics of networks of neurons that obey membrane equations. The resulting architecture is structurally similar to the mammalian retina, requiring a mechanism for temporal adaptation analogous to photoreceptors, followed by cells of opposite polarity analogous to on and off bipolar cells, and finally a layer of ganglion cells that summate bipolar cell inputs. The model predicts that X and Y cells consist of the same neural mechanism acting in different parametric regimes. In agreement with morphological and physiological data, analyses and numerical simulations show that an increase in receptive field center size changes the model’s response from X-like to Y-like. This functional duality results from mathematical properties of the push-pull shunting network, which can selectively enhance sustained or transient response components on the basis of RF morphology. The model also explains how it is possible for X cells to respond linearly to a broad range of spatial and temporal modulation functions in spite of arbitrary nonlinear preprocessing. Finally, in general agreement with biological data, the model predicts that stationary images lead to loss of instantaneous contrast at equilibrium, i.e., that stabilized images fade, as an unavoidable side effect of photoreceptor adaptation.