Recurrent Long - Range Interactions in Early Vision 1

1 Motivation: Functionality and Architecture How does the brain manage to form invariant representations of the environment that are relevant for the current behavioral task? The sensory system is steadily confronted with a massive information ow that arrives via diierent channels. In vision, spatio-temporal pattern arrangements that signal coherent surface arrangements must be somehow reliably detected and grouped into elementary items even in changing situations and under variable environmental conditions. Such a grouping enables the segregation of gural components from cluttered background as well as the adaptive focussing of processing capacities while suppressing parts of the scene that are less relevant to support the behavioral goal or task ((4, 8]). Grouping and segregation requires the interaction of several representations and activity distributions generated by diierent processing streams. Here we focus on the detection of contour features such as smooth boundary patterns as well as corner and junction conngurations by adaptive neural mechanisms. A characteristic feature of cortical architecture is that the majority of visual cortical areas are linked bidirectionally by feedforward and feedback bre projections to form cortico-cortical loops. So far, the precise computational role of the descending feedback pathways at diierent processing stages remains largely unknown. Empirical evidence suggests that top-down projections primarily serve to modulate the responsiveness of cells at previous stages of the processing hierarchy (e.g. 7]). We particularly investigated the recurrent interaction of areas V1 and V2. The results of this investigation suggest a novel interpretation of the role of contour grouping and subjective contour interpolation at V2 such that observable eeects relate to the task of surface segmentation. This information is used to evaluate and selectively enhance initial measurements at the earlier stage of V1 processing of oriented contrasts. Other architectural principles encountered in cortical architecture are long-range horizontal connections and intra-cortical feedback loops ((3]), among others. Via horizontal connections cells of like-orientation couple and thus cell responses are selectively innuenced by stimuli outside their classical receptive eld (RF). We propose a simpliied model architecture of V1 that incorporates a sequence of preprocessing stages and a recurrent loop based on long-range interaction. The results demonstrate that noisy low contrast arrangements can be signiicantly enhanced to form elementary items of smooth contour segments which are precursory for subsequent integration and organization into salient structure. Beyond the formation of salient contour fragments this scheme of processing is able to enhance contour responses at corner and junction conngurations. …