Computer Science Colloquium

Domain Adaptation with Multiple Sources

Yishay Mansour

Friday, November 7, 2008 11:30 A.M.
Room 1302 Warren Weaver Hall
251 Mercer Street
New York, NY 10012-1110

Directions: http://cs.nyu.edu/csweb/Location/directions.html
Colloquium Information: http://cs.nyu.edu/csweb/Calendar/colloquium/index.html

Host

Mehryar Mohri mohri@cs.nyu.edu, (212) 998-3200

Synopsis

We will discuss problems that arise in machine learning regarding domain adaptation. The motivation for the domain adaptation problem is to be able to generalize from domains which are given, to new domains. The practical motivation is to avoid the need to re-learn from scratch a good hypothesis for each domain.

In this talk we will concentrate on a theoretical analysis on a very specific domain adaptation problem with multiple sources, specified as follows. For each source domain, the distribution over the input points as well as a hypothesis with error at most epsilon are given. Our problem consists of combining these hypotheses to derive a hypothesis with small error with respect to the target domain. We present several theoretical results relating to this problem. In particular, we prove that standard convex combinations of the source hypotheses may in fact perform very poorly and that, instead, combinations weighted by the source distributions benefit from favorable theoretical guarantees. Our main result shows that, remarkably, for any fixed target function, there exists a distribution weighted combining rule that has a loss of at most epsilon with respect to any target mixture of the source distributions. We further generalize the setting from a single target function to multiple consistent target functions and show the existence of a combining rule with error at most 3*epsilon. Finally, we report some empirical results for a multiple source adaptation problem with a real-world dataset.

This is joint work with M. Mohri and A. Rostamizadeh.

Bio

Prof. Yishay Mansour has done his PhD studies at MIT, performed a postdoctoral in Harvard and worked in IBM T. J. Watson Research Center. Since 1992 he is at Tel-Aviv University, where he is currently a Professor of Computer Science and has serves as the head of the School of Computer Science during 2000-2002. Prof. Mansour has held visiting positions with Bell Labs, AT&T research Labs, and IBM. Currently he is visiting Google Inc. in New York.

Prof. Mansour has published over 50 journal papers and over 100 proceeding paper in many various areas of computer science with special emphasis on communication networks and machine learning. Prof. Mansour is currently an associate editor in a number of distinguished journals (ACM Transactions on Algorithms, Machine Learning Journal, Journal of Machine Learning Research, Mathematics of Operations Research, International Journal of Game Theory, and Siam J. on Computing) and has been on numerous conference program committees. He was both the program chair of COLT (1998) and served on the COLT steering committee. He has supervised over a dozen graduate students in various areas including communication networks, machine learning and algorithm design.


top | contact webmaster@cs.nyu.edu