Abstract

A neurallike model (a Boltzmann machine) was built to extract the true 2-D motion of an entire pattern from ambiguous local motion information available at the pattern’s component contours (i.e., it solves the aperture problem). The model has an input and output layer representing visual cortical areas V1 and MT, respectively. Area MT, an area involved in motion analysis, receives a direct topographic projection from V1. V1 neurons act as local motion detectors in that they can only measure the component of motion perpendicular to the orientation of a moving contour. In contrast, ~20% of MT neurons possess pattern direction selectivity—selectivity for the motion of a pattern as a whole. In the model, each unit is selective for a specific speed and direction of motion. Connectivity is restricted so that each V1 unit projects only to the MT units with which it is consistent (i.e., that could describe its true motion). The input to the system is the set of velocity vectors perpendicular to the contours of a moving pattern. Input units are clamped to represent the input, and the system is allowed to relax to find the state which represents the minimum energy configuration given the input. The network relaxes quickly and the MT units describing the motion of the entire pattern are maximally activated.

© 1986 Optical Society of America