Neural Principles of Preattentive Scene Segmentation: Hints from Cortical Recordings, Related Models, and Perception

Abstract

Preattentive segmentation of visual scenes is a prerequisite of object recognition and effective visuomotor coordination. For this the visual system has to specify neural representations of contours and regions of potential relevance so that top-down acting mechanisms of attention, expectation and visual memory can interact with them. This chapter attempts to show how the largely unknown neural mechanisms of scene segmentation may be uncovered. Starting from principles of perceptual grouping, we present experimental results mainly of our own multiple microelectrode recordings from the visual cortex of awake monkeys. Perceptual and experimental hints at principles of scene segmentation are supported by related simulations of spike coding networks of minimal complexity. We follow the hypothesis that fast signal coupling and decoupling among visual cortical circuits define preattentively relations among scene segments. Two properties of scene segments and related signal coupling are differentiated: (1) Transient retinal changes evoke short simultaneous activations that are coarsely synchronized across the entire representation of the changing segments; (2) During ocular fixation stable retinal images induce fast cortical oscillations (FCOs; 30 – 90 Hz) that are phase-correlated along the cortical representation of contours and across segment regions. Phase coupling among FCOs in neighboring cortical populations is weak, distributions of phase differences are symmetrical to zero delay and their width increases with cortical distance. These properties explain the small size of cortical patches over which coherent FCOs have previously been reported. The representation of such a cortical patch in visual space (defined by the superimposed classical receptive fields (cRFs) of the synchronized neurons) is termed here the feature association field (AF). Arguments from experiments and simulations are developed showing that AF size at a lower level of processing can explain the larger cRFs at the next level and hence a stepwise increase in establishing relevant scene relations. In the light of our new data, the previous hypothesis of feature association by synchronization is modified to a hypthesis of coding region, contour and object continuity by phase-continuity, including single event stimulus-locked and rhythmic FCO processes. (This article contains some new unpublished experimental and modeling results.)