Leslie Valiant
Division of Engineering and Applied Sciences
Harvard University
Monday, November 17, 1997
11:00 AM
McDonnell Douglas Auditorium, UCI
By cognitive computation we mean here a combination of the basic memory, learning and reasoning tasks that are needed to build, manipulate and use large amounts of knowledge about a complex environment. The domains of interest are those that are too complex to be systematized as exact sciences, and are typified by those that are studied in AI in the context of commonsense reasoning.
Research on commonsense reasoning over several decades has emphasized programmed systems, as opposed to those that learn. A number of fundamental issues have been identified that present at least potential difficulties to the programmed system approach. These include the problem of incomplete information, nonmonotonic phenomena, robustness to errors, inconsistencies in the knowledge, the issue of relevance, and conflict resolution. On top of all this is the problem of formulating all proposed solutions by mechanisms having feasibly low computational complexity.
In this talk we will discuss a proposed architecture that puts massive learning, as opposed to programming, at the center. We suggest that recent work on the theory and practice of machine learning and on abstract models of neural computation has uncovered a set of mechanisms that can be brought together in this one architecture to provide a convincing alternative approach. We shall argue that this approach solves in principle some of the old problems. The new problem area that emerges is that of generating suitable teaching materials for such systems.
Distinguished Speaker Series Home Page
Information and Computer Science
University of California, Irvine
Irvine, CA 92697-3425
September 19, 1997