High-Level Vision: Top-Down Processing in Neurally Inspired Architectures

摘要

An exhaustive literature review of computationally relevant studies of high-level vision and mental imagery was conducted, and a qualitative theory of the processing subsystems in the brain and their interactions during visual object identification are summarized herein. A computational model that embodies these ideas was built with the intention that it could be 'damaged' in various ways in order to observe its behavior while performing visual tasks that are analogous to those performed by the brain-damaged patients; this model is described in detail in this report. A second model of early/intermediate vision, which is intended to supplement and complement the work on late visual processing, has also been developed, and is described. Progress testing the models is reported. Additional deliverables include a binder with the abstracts of all the relevant literature (which can serve as a guide for further research) and an annotated copy of the software code for the IMPER model.