Workshop on Advances in Machine Learning

Montreal, Canada, June 8-11, 2003


ORGANIZERS: Yoshua Bengio, Balázs Kégl (University of Montreal) Doina Precup (McGill University)

SPONSORS: Centre de Recherches Mathématiques, MITACS,



Probabilities are at the core of recent advances in the theory and practice of machine learning algorithms. The workshop will focus on three broad areas where these advances are crucial: statistical learning theory, learning algorithms, and reinforcement learning. The workshop will therefore bring together experts from each of these three important domains. Among the sub-topics that will be covered, we note: variational methods, graphical models, the curse of dimensionality, empirical methods to take advantage of theories of generalization error, and some of the applications of these new methods.

On the theoretical side, in recent years a lot of effort has been devoted to explain the generalization abilities of popular learning algorithms such as voting classifiers and kernel methods. Some of these results have given rise to general principles that can guide practical classifier design. Some (non-exclusive) sub-topics in this aspect of the workshop include Rademacher and Gaussian complexities, algorithmic stability and generalization, localized complexities and results on the generalization ability of voting classifiers and kernel-based methods.

On the algorithmic side, one of the emphasis of recent years has been on probabilistic models that attempt to capture the complex structure in the data, often by discovering the main lower-dimensional features that explain the data. This raises interesting and difficult questions on how to train such models, but such algorithms may have wide ranging applications in domains in which the data has interesting structure that may be explained at multiple levels, such as in vision and language.

In reinforcement learning (RL), recent research has brought significant advances in some of the traditional problems, such as understanding the interplay between RL algorithms and function approximation, and extending RL beyond MDPs. At the same time, new areas of research, such as computational game theory, have developed at the interface between RL and probabilistic learning methods. In this workshop, we invite presentations on all RL topics, ranging from theoretical development to practical applications.


Satinder Singh, U. Michigan, Ann Arbour, MI, USA
Sridhar Mahadevan, U. Massachusetts, USA
Michael Littman, Rutgers U., NJ, USA
Peter Bartlett, U. California Berkeley, CA, USA
Vladimir Koltchinskii, U. New Mexico, NM, USA
Imre Risi Kondor, Columbia U., USA
Yann Le Cun, NEC Research, NJ, USA
Paolo Frasconi, U. Firenze, Italy
Dale Schuurmans, Waterloo U., Ontario, Canada
Sam Roweis, U. Toronto, Ontario, Canada
Geoff Hinton, U. Toronto, Ontario, Canada


There will be no paper proceedings, but with the authorization of authors we will post the presented papers on the web.
Since the accepted papers will not be published, they can of course be submitted to other venues.


The registration fees are minimal: regular registration fees are 100$CAN. Reduced rate for students from a Canadian academic institution: 50$CAN. Follow this link to register. Please note that the registration software is optimized for Netscape and Internet Explorer, and it might not work with other browsers.


The workshop will take place at the Centre de Recherches Mathematiques, on the campus of Universite de Montreal, in lively and beautiful Montreal, Canada. The conference will be held in the Pavillon Andre Aisenstadt, 2920 chemin de la Tour.


Follow the registration link to have accomodations booked for you or see a list of suggested hotels. Note that you will probably have to plan your return from Montreal on or before the 12th of June.