Effects of the ganglion cell response nonlinear mapping on visual system’s noise filtering characteristics

Abstract

Up to the cortical simple cells there exist some good linear models of the visual system neurons’ dynamic receptive fields. These linear approximations are based on considerable amount of experimental evidences. Based on the linear approach, assumptions can be made about noise filtering characteristics of the early visual processing. Unfortunately, these notions, based primarily on power spectral estimations, can not be used when neurons’ responses reach the nonlinear part of their dynamic range. Real neurons’ response, however, can frequently reach this region. We made simple Cellular Neural Network (CNN) models to simulate the retinal ganglion cells dynamic receptive fields. Using an apt inversion method, in the root mean square error sense optimal reconstruction can be achieved even from the modeled nonlinearly mapped responses. This way we can predict the nonlinear system’s noise filtering properties. Using the CNN as a modeling frame it is easy to implement both the deconvolution and the necessary additional processing steps as well to establish reconstruction. By this technique, depending on the properties of the additive and intrinsic noise terms we could estimate, from the noise filtering point of view, ideal parameters of the dynamic receptive fields’ linear and even simple nonlinear functions. Our results can explain the measured effects of dark adaptation on the receptive field structure and can give some insight to the design of the probable further information processing steps. This type of preprocessing can ameliorate the efficiency some of the existing image compression algorithms, and using CNN technology, the necessary reconstruction can be even accomplished in real time.