Robustness from Non-Stop Collaboration

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

Proposal 0534978 "Robustness from Non-Stop Collaboration" PI: Patrick H. Winston MIT This proposal seeks to understand how robust behavior can emerge from modules that collectively develop descriptions and cooperatively solve problems at a level beyond the ability of any one module operating individually. For instance, human language and vision systems cooperate to form a world view that is beyond what either could produce individually. A key aspect of this project is to understand how the necessary intermodular communication capabilities grow as the modules grow in their individual capabilities, and how such communication can be tolerant of imperfections in the communications. This research approaches this problem through development of composable intermodular channels, using two enabling ideas: fusion maps and self-configuring communication channels. Fusion maps use constraint propagation, elevated to operate between modules and across representations, to determine semantic correspondences between modules and to support bi-directional transfer of knowledge from a module that is more certain of information to those that are less so. Self-configuring communication channels learn constraints between modules from shared experiences; this idea addresses the twin problems of when should a channel attempt to learn and how can useful equivalences actually be learned. The project will investigate these ideas on the problem of forming a composite record of activities from data collected from a diverse set of sensors (e.g., images, sound), for instance, in an office domain where there can be a variety of activities. Besides giving insight into the nature of intelligence, this research will improve the robustness of computer systems, important in many contexts from the control of dangerous equipment to the mitigation of natural disasters.