CONNECTIONIST MODELS AND PARALLELISM IN HIGH-LEVEL VISION

Students of human and machine vision share the belief that massively parallel processing characterizes early vision. For higher levels of visual organization, considerably less is known and there is much less agreement about the best computational view of the processing. This paper lays out a computational framework in which all levels of vision can be naturally carried out in highly parallel fashion. One key is the representation of all visual information needed for high level processing as discrete parameter values which can be represented by units. Two problems that appear to require sequential attention are described and their solutions within the basically parallel structure are presented. Some simple program results are included.