Lateral Interactions and Feedback

Abstract

So far, we have almost exclusively considered a “bottom-up” or feedforward framework, in which the incoming image is processed in a number of successive stages, the information flowing in one direction only. However, it is widely appreciated in visual neuroscience that the brain is doing something much more complicated than just feedforward processing. There is a lot of evidence for 1. feedback from “higher” areas to lower areas, e.g., “top-down” connections from V2 back to V1, as well as 2. lateral (horizontal) interactions, by which we mean here connections between features in the same stage, e.g., connections between simple cells. In this chapter, we will see how such phenomena are rather natural consequences of Bayesian inference in the models we have introduced. First, we will introduce a model of feedback based on thresholding, or shrinkage, of coefficients in the higher stage. Second, we will consider a lateral interaction phenomenon: end-stopping in simple cells. Finally, we will discuss the relationship of the principle of predictive coding to these phenomena.