Due to travel by some of us to a DARPA-related meeting. The tea talk this week has been moved to Friday. Hope you all can make it -- this talk should be of interest to everyone and particularly for the newer students that may not have had much previous exposure to statistical learning theory. The talk will be given by Ian Goodfellow.

Date/Time: Friday Nov. 4th / 14h00
Location: LISA Lab (AA3256)

Title: Statistical Learning Theory

Abstract:
I will give a tutorial on statistical learning theory. First I will
show how to prove some simple probabilistic guarantees on the
generalization error of a classifier in the PAC learning framework
using Chebyshev's inequality. Next I will discuss how to extend these
bounds to more realistic settings using the concept of
Vapnik-Chervonenkis dimension. Finally I will compare structural risk
minimization, a model selection principle based on statistical
learning theory, to other methods of obtaining low generalization
error.


Cheers,
Aaron

--
Aaron C. Courville
Département d’Informatique et
de recherche opérationnelle
Université de Montréal
email:Aaron.Courville@gmail.com