CRM/ MITACS/ IRIS Workshop on Selecting and Combining Models with Machine Learning Algorithms

Montreal, Québec, Canada, April 11-14, 2000

April 11-14, 2000
Centre de Recherches Mathématiques, Montréal

Organizers: Yoshua Bengio (Université de Montréal) and Dale Schuurmans (University of Waterloo)

Workshop Program and Pointers to Workshop Material


Location of the conferences: Room 1140 (1st floor), Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour, Montréal

 

Workshop Objectives

A central objective of machine learning research is to develop algorithms that learn predictive relationships from data. This is a central component of data mining and knowledge discovery tasks, which are becoming commonplace applications in the realm of e-commerce. This is a difficult task, however, because inferring a predictive function from data is in fact an "ill-posed" problem; that is, many functions can often "fit" a given finite data set, and yet these functions might generalize very differently on new data drawn from the same distribution.
To make this problem well-posed one needs to somehow "calibrate" the complexity of the proposed function class to the amount and quality of available sample data. A classical approach is to perform "model selection" where one imposes a preference structure over function classes and then optimizes a combined objective of class preference and data fit. In doing so, however, it would be useful to have an accurate estimate of the expected generalization performance at each preference level; one could then pick the function class that obtained the lowest expected error, or combine functions from the functions classes with the lowest expected error, and so on. Many approaches have been proposed in the past for this purpose, both in the statistics and the machine learning research communities.
Recently in machine learning there has been significant interest in new techniques for evaluating generalization error, for optimizing generalization error, and for combining and selecting models. This is exemplified, for example, by recent work on Structural Risk Minimization, Support Vector Machines, various Boosting algorithms, and the Bagging algorithm. These new approaches suggest that better generalization performance can be obtained using new, broadly applicable procedures. Progress in this area has not only been important for improving our understanding of how machine learning algorithms generalize, it has already been demonstrated to be very useful for practical applications of machine learning and data analysis.
This workshop will bring together several key researchers in the fields of machine learning and statistics to present their recent results and debate the controversial issues that have been dividing them in the recent machine learning and neural network conferences. The following leaders in this field have accepted to participate to the workshop as invited speakers:

Peter Bartlett (Australia National University), Leo Breiman (University of California-Berkeley), Tom Dietterich (Oregon State University), Yoav Freund (AT&T Labs-Research), Michael Perrone (IBM T.J. Watson Research Center), Robert Schapire (AT&T Labs-Research), Peter Sollich (King's College, University of London), Grace Wahba (University of Wisconsin at Madison).

The workshop is sponsored by the CRM (Centre de Recherches Mathématiques) as well as by the MITACS (Mathematics of Information Technology And Complex Systems) and IRIS (Institute for Robotics and Intelligent Systems) Networks of Centers of Excellence.

Registration and Accomodation

There are no registration fees but participants are urged to register in advance as the number of attendees may be limited. Follow this link for registration and accomodation information: REGISTRATION PAGE