[Colloquium] Guest Speaker

[Ponda Barnes]

Guest Speaker
 
Presented by: Toyota Technological Institute at Chicago
 
Speaker: Ambuj Tewari
Speaker's home page: http://www.eecsberkeley.edu/~ambuj/
 
Date: Tuesday, April 10, 2007
Time: 10:00
Location: TTI-C Conference room
 
Title: Consistency of Multiclass Classification Methods (Joint work with
Peter Bartlett)
 
Abstract: 
 
Many important binary classification methods, such as SVMs and AdaBoost,
involve solving a convex minimization problem based on a loss function that
acts as a surrogate for the 0-1 loss.  Statistical consequences, especially
consistency, of the choice of the loss function have been thoroughly
analyzed in the literature.  The theory is now being extended to multiclass
classification. In this talk, I will describe a geometric way of looking at
the problem of consistency of multiclass methods. This will lead us to a
necessary and sufficient condition characterizing consistency of a large
family of multiclass methods.  The results will then be applied to answer
consistency questions for several methods already proposed in the
literature.  Along the way, we will see why multiclass consistency is a more
subtle issue than its binary counterpart. 
That will help us understand why seemingly straightforward generalizations
of consistent binary methods do not always preserve consistency.
 
If you have any questions or would like to meet the speaker, please contact
Ponda Barnes at pondabarnes at tti-c.org
For future TTI-C talks and events please go to
http://ttic.uchicago.edu/cal/month.php
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20070409/994d5fd4/attachment.htm