Recurrent Long-Range Interactions in Early Vision

Abstract

A general principle of cortical architecture is the bidirec- tional flow of information along feedforward and feedback connections. In the feedforward path, converging connections mainly define the fea- ture detection characteristics of cells. The computational role of feedback connections, on the contrary, is largely unknown. Based on empirical find- ings we suggest that top-down feedback projections modulate activity of target cells in a context dependent manner. The context is represented by the spatial extension and direction of long-range connections. In this scheme, bottom-up activity which is consistent in a more global context is enhanced, inconsistent activity is suppressed. We present two instantia- tions of this general scheme having complementary functionality, namely a model of cortico-cortical V1-V2 interactions and a model of recurrent intracortical V1 interactions. The models both have long-range interac- tions for the representation of contour shapes and modulating feedback in common. They differ in their response properties to illusory contours and corners, and in the details of computing the bipole filter which models the long-range connections. We demonstrate that the models are capable of basic processing tasks in vision, such as, e.g., contour enhancement, noise suppression and corner detection. Also, a variety of perceptual phe- such as grouping of fragmented shape outline and interpolation nomena of illusory contours can be explained.