The Role of Simplicity in Learning Theory
Date and Time
Wednesday, February 6, 2002 - 4:00pm to 5:30pm
Location
Computer Science Small Auditorium (Room 105)
Type
Colloquium
Speaker
David McAllester, from AT&T Research
Host
Andrew Appel
Science values simplicity. All other things being equal, a
simple explanation is preferable to a complex one. Bayesians assign
higher prior probability to simple theories. But is it really true
that a simple theory is a-priori more likely than a complex one? It
turns out that one can justify a preference for simplicity independent
of Bayesian assumptions. The justification involves only the law of
large number and the observation that the number of simple theories is
limited. This talk will present this justification and go on to
describe more general "laws of large numbers" that justify more
sophisticated methods of evaluating the accuracy of predictive rules.