|
|
|
 |
|
Department of Statistics, University of California, Berkeley, USA
Wednesday 24 February 2010
16.00
Seminar Room B10 (Basement)
Alexandra House, 17 Queen Square, London, WC1N 3AR
Group Lasso extensions and sparse structured dictionary learning
The Group Lasso as been proposed as a generalization of the Lasso to the case where variables are selected in groups. However, both the theory and the algorithms assume that the groups are disjoint. In this talk, I will show how considering overlapping groups leads to two types of variable selection problems. In each case, it is possible to construct a convex regularization which generalizes the Group Lasso, with efficient algorithms to solve the corresponding optimization problems. I'll show an application to "pathway selection" in genomics and how such regularizations can be used in the context of dictionary learning to
learn structured dictionaries. In particular, I'll show how an appropriate choice of regularization allows to estimate hierarchical topic models for textual data comparable to those obtained using non-parametric Bayesian methods.