Abstract
Natural scenes are known to have a variety of similar statistical properties. In particular, scenes are found to have Fourier power spectra that fall off with frequency as 1/frequency^2 (amplitude falls as 1/f). One popular theory of retinal ganglion cell processing suggests that the goal of this unit is to provide a decorrelated or “whitened” output given natural stimuli. That is, by virtue of their center-surround receptive field organization, ganglion cells exploit the predictable statistics in scenes in order to represent the input more efficiently. But there are a number of spatial decorrelation strategies that do not require a receptive field organization like that found in primate retinal ganglion cells. In this study, we compare different decorrelation strategies for natural scene stimuli to determine the relative advantages of center-surround organization. We show that the localized center-surround organization produces more than decorrelation: it also gives a more sparse response. In addition, we argue that the relative sensitivity of ganglion cells as a function of size—for at least P-type ganglion cells—supports the notion that these cells achieve some degree of response equalization when given natural scenes. We also consider the possible role played by the neurons' known nonlinearities in producing a more efficient response.