The UdeM-McGill-MITACS machine learning seminar series has its last Fall seminar this Friday. Please note the changed location and time:

Label Embedding Trees for Large Multi-Class Tasks

by Samy Bengio 
Google Research

Location: Pavillon Claire-McNicoll (UdeM), room Z-209
TimeFriday, December 17, 10:30 (!)

Abstract
Multi-class classification becomes challenging at test time when the number of classes is very large and testing against every possible class can become computationally infeasible. This problem can be alleviated by imposing (or learning) a structure over the set of classes. We propose an algorithm for learning a tree-structure of classifiers which, by optimizing the overall tree loss, provides superior accuracy to existing tree labeling methods. We also propose a method that learns to embed labels in a low dimensional space that is faster than non-embedding approaches and has superior accuracy to existing embedding approaches.
Finally we combine the two ideas resulting in the {\em label embedding tree} that outperforms alternative methods including  One-vs-Rest while  being  orders of magnitude faster.