TECHNIQUES FOR CONSTRUCTING AN ASYMPTOTICALLY TIGHT LEARNING THEORY
Professor John Langford, Department of Computer Science, Carnegie Mellon University

Abstract: Learning theory, while qualitatively useful, has suffered from an inability to tightly bound the true error rate of a learned classifier. This statement is especially strong for classifiers drawn from continuous spaces. I will discuss techniques for achieving asymptotically tight bounds on the future error rate of a classifier learned on a continuous space. Proofs will be avoided in the interest of clarity, and I will present empirical results showing nonvacuous bounds with just 100 examples on neural networks.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 13 July 2001.

back