What do statistical models of natural images tell us about visual processing in the brain?
Yan Karklin, Carnegie Mellon University
How does our visual system represent the continuum of complex patterns, shapes, and textures we encounter in the visual world? One approach to answering these questions is to understand the statistical regularities underlying these complex patterns — if the goal of the early visual system is to efficiently encode its input, then it should be tuned to the statistical structure of the input, and a good model of this structure will provide insight into neural computation. Early models of natural scene statistics helped explain the linear properties of neurons in the primary visual cortex. However, even neurons early in the visual pathway exhibit strongly non-linear behavior, and neurons in higher-order visual areas cannot be characterized by linear models at all. We propose a hierarchical statistical model that naturally extends a simple multi-variate Gaussian model and captures higher-order dependencies in the data. The model makes few assumptions about the structure of the data or the types of computations performed by the visual system, and makes interesting predictions about complex adaptive behavior.
Abstract Author(s): Yan Karklin and Michael S. Lewicki