Contrast Gain Control in Natural Image Representations

The Redundancy Reduction principle by Barlow and Attneave has been a very influential idea for the understanding of neural representations in the early visual system. When applied to ensembles of natural images in linear image models several features emerge that are strikingly similar to features of the early visual system. However, several recent findings indicate that the linear framework is too limited to demonstrate a clear advantage (i) in terms of redundancy reduction or (ii) for higher visual tasks like e.g. object recognition. Contrast gain control or divisive normalization is an ubiquitous nonlinear neural response property that has been been shown to reduce higher order correlations in natural images and to be an important preprocessing step in vision algorithms. In this talk, I will present our recent work on the quantitative evaluation of contrast gain control for modelling and representing natural images. To this end we use the class of Lp-spherically symmetric distribution and our recent generalization, the class of Lp-nested symmetric distributions, to model the local statistics of natural images patches. Our main findings are that (i) those models perform better than other density models like Independent Component Analysis (ICA) or Independent Subspace Analysis (ISA) that have been used in the context of early vision before and (ii) that contrast gain control is of much greater importance for redundancy reduction than the actual shape of the linear filters. Together with recent results by Jarrett et al. (ICCV'09) our findings elaborate on the fact that nonlinear mechanisms are crucial for the design of powerful representation of natural images