Probabilistic models of early vision

Abstract

How do our brains transform patterns of light striking the retina into useful knowledge about objects and events of the external world? Thanks to intense research into the mechanisms of vision, much is now known about this process. However, we do not yet have anything close to a complete picture, and many questions remain unanswered. In addition to its clinical relevance and purely academic significance, research on vision is important because a thorough understanding of biological vision would probably help solve many major problems in computer vision.

A major framework for investigating the computational basis of vision is what might be called the probabilistic view of vision. This approach emphasizes the general importance of uncertainty and probabilities in perception and, in particular, suggests that perception is tightly linked to the statistical structure of the natural environment. This thesis investigates this link by building statistical models of natural images, and relating these to what is known of the information processing performed by the early stages of the primate visual system.

Recently, it was suggested that the response properties of simple cells in the primary visual cortex could be interpreted as the result of the cells performing an independent component analysis of the natural visual sensory input. This thesis provides some further support for that proposal, and, more importantly, extends the theory to also account for complex cell properties and the columnar organization of the primary visual cortex. Finally, the application of these methods to predicting neural response properties further along the visual pathway is considered.

Although the models considered account for only a relatively small part of known facts concerning early visual information processing, it is nonetheless a rather impressive amount considering the simplicity of the models. This is encouraging, and suggests that many of the intricacies of visual information processing might be understood using fairly simple probabilistic models of natural sensory input.