Abstract
It is proposed that the early visual system exploits the statistical structures of the visual environment in order to represent the visual input efficiently. In previous studies, it has been shown that the efficient representations of natural images are localized and oriented filters similar to the receptive fields of simple cells in the visual cortex. However, the orientation selective receptive fields do not emerge before visual cortex, and a simple cell receives inputs from both ON and OFF ganglion cells. Because the ON and OFF ganglion cells process visual input in a nonlinear way, one cannot simply study the efficient coding as a linear process. In the current research, the effect of static nonlinearity on the efficient coding of the visual input is investigated. Natural time varying images are preprocessed with a biologically inspired center-surround filter (CSF). Similar to the earlier studies, the efficient representations of the direct CSF output are localized and oriented filters. However, the efficient coding of the rectified CSF output (ON/OFF channels) does not result in those filters. Instead, the filters have center surround structures similar to those of ganglion cells and most of the improvement in efficiency results from the rectification. Furthermore, the efficiency can be improved much more by using a temporal filter similar to the temporal receptive field of lateral geniculate nucleus (LGN) cells. In conclusion, the results suggest that using biologically inspired spatial/temporal filter of retina/LGN with static nonlinearity gives more efficient representation than linear processing.