Temporal information as top-down context in binocular disparity detection

Abstract

Recently, it has been shown that motor initiated context through top-down connections boosts the performance of network models in object recognition applications. Moreover, models of the 6-layer architecture of the laminar cortex have been shown to have computational advantage over single-layer models of the cortex. In this study, we present a temporal model of the laminar cortex that applies expectation feedback signals as top-down temporal context in a binocular network supervised to learn disparities. The work reported here shows that the 6-layer architecture drastically reduced the disparity detection error by as much as 7 times with context enabled. Top-down context reduced the error by a factor of 2 in the same 6-layer architecture. For the first time, an end-to-end model inspired by the 6-layer architecture with emergent binocular representation has reached a sub-pixel accuracy in the challenging problem of binocular disparity detection from natural images. In addition, our model demonstrates biologically-plausible gradually changing topographic maps; the representation of disparity sensitivity changes smoothly along the cortex.