Computational models of visual processing: a discussion of normalization and predictive coding using Marr's three levels of analysis

Abstract

Our percept is the result of processing visual information by our visual system. This system contains many billions of neurons and a manyfold of connections between these neurons. In order to understand a complex system like our visual system, it has been proposed to approach it at different levels of analysis. This proposal distinguishes three levels: the first level describes the computational goal of the system, the second level the algorithms the system uses to achieve this goal, and the third level describes the way in which these algorithms are physically implemented in the brain. Focusing mainly on the first two levels of analysis, this review describes two computational models: the normalization model and the predictive coding model. This discussion shows that, despite little is known about the biological mechanisms underlying the models, the behavior of the visual system can be very well understood when computational models at the remaining two levels of analysis are applied to its behavior. Furthermore, this discussion shows that the models' algorithms may supplement one another at certain stages of visual processing. Finally, this review may provide support for a relation between predictive coding and normalization models. Several implications of this possible relation are discussed.