Computational and Psychophysical Study of Human Vision Using Neural Networks
Abstract
The overall goal of our research program is to construct models of the human visual system that can be implemented on available computers and capture essential abilities of the real thing. These models should be useful in understanding how the human visual system works and for practical applications. In order to incorporate some of the known structural features of the brain in our models, we have chosen a neural net paradigm to mimic some aspects of the real nervous system. These networks contain nodes representing simplified nerve cells and can have an enormous variety of structures, some of which are the subjects of intensive study in many laboratories. Since so many different network structures are possible, it is necessary to use as much information as possible to limit the choice of nets to those most likely to be useful models of the human visual system. Our work in psychophysics is designed to provide limits on the choice of architectures for model nets by requiring them to satisfy certain general conditions indicated by these experiments. Several experimental projects will be described below concerning perception of relative depth and motion. One generalization that emerges from all of them is that local visual judgements can be grossly influenced by information gleaned from quite distant parts of a scene. To mimic the operation of the human visual system, then, a neural net must collect information from sizeable areas of a scene and use it to influence outputs from local visual processes. (SDW)
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 28, 1989
- Accession Number
- ADA213290
Entities
People
- Donald A. Glaser
- Kumar Tribhawan
Organizations
- University of California, Berkeley