Department of Mathematics

Speaker: James A. Moorer
Title: Bayes Rules! Statistical Engines Mimic Intelligent Behavior
Affiliation: Adobe Systems, Inc.
Date: Friday, October 28, 2011
Place and Time: Room 101, Love Building, 3:35-4:30 pm
Refreshments: Room 204, Love Building, 3:00 pm

Abstract. The debut of IBM's Watson was a triumph of what might be termed "dumb" statistics. Despite having no facility for grammar or syntax, Watson could produce super-human behavior just by statistical association of words over a very large training corpus, albeit in a very limited and well-defined domain. Watson is one example among many that are appearing in popular technology as varied as smart phones, digital cameras, web searches, speech recognition, and more that share a common computational architecture: large numbers of relatively simple operators are used to produce "features" that are then forwarded to statistical discrimination techniques and subjected to intense training. The performance of the resulting systems is startling when we consider the paucity of the information content of the features these systems use. It also raises some questions on the nature of intelligence and what we mean by intelligent behavior. The talk will present a number of examples of this kind of system in use or under development today with emphasis on some of the mathematics involved.