Vision, Instruction and Action
Abstract
This thesis describes Sonja, a system which uses instructions in the course of visually-guided activity. The thesis explores an integration of research in vision, activity, and natural language pragmatics. Integrated systems are uncommon in Artificial Intelligence research, but they provide otherwise unobtainable insights into the relationships between AI problems and techniques. The design of Sonja is motivated by psychophysical and neuroscientific evidence. The computations the system performs are compatible with the constraints imposed by neurally plausible hardware. The system has access to its domain only via the domain's primitive actions and via a simulated visual system. This visual system addresses the fundamental visual problems of selectively applying visual processing to subsets of the image, finding regions of the image with task-relevant properties, and establishing spatial relationships among parts of the image. It unifies and provides the first psychophysically accurate implementations of a variety of intermediate visual processes, particularly visual search and routines. It models previously unspecified aspects of these phenomena and demonstrates by connecting proposed mechanisms with a natural task domain that they are in fact useful. (kr)
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1990
- Accession Number
- ADA228626
Entities
People
- David Chapman
Organizations
- Massachusetts Institute of Technology