Friday, April 06, 2007
12:30 PM
CSB 703
Nick Cassimatis
University of Rochester
A Cognitive Substrate for Human-Level Intelligence
People can reason, talk and learn about a wide variety of topics (from simple everyday physical events to parliaments, airplanes and interest rates) . How is this possible given that our minds evolved while most of these did not exist? How does the mind manage to deal with situations that can involve many, often unanticipated, combinations of domains? In this talk, I will address these questions with the Cognitive Substrate Hypothesis. It states that a relatively small set of abilities -- a preliminary list includes reasoning about objects, relations, locations, temporal intervals, categories, and parts -- is sufficient to produce human-level reasoning in most or all domains. Using a model of elementary physical reasoning as a substrate, our lab has modeled reasoning about natural language syntax (using HPSG) and reasoning about people's beliefs and desires. These reductions allowed reasoning in each domain to be constrained by reasoning in another. For example, our English parser is able to use knowledge about objects' categories and people's intentions to resolve ambiguities simply (and without specifically providing for this functionality) as a consequence of representing and reasoning about syntax, objects and intentions within the same substrate. This work provides evidence for the cognitive substrate hypothesis and demonstrates that an implementation of a cognitive substrate makes building intelligent systems and cognitive models in many domains a (relatively) simple matter of mapping that domain onto the cognitive substrate.