Seminar Details
2014-02-11 (14.00) : Convex approaches to structured sparsity
At Euler building
Organized by Mathematical Engineering
Speaker :
Guillaume Obozinski (Ecole des Ponts - ParisTech)
Abstract :
Sparse estimation methods are aimed at using or obtaining parsimonious representations of data or models. They were first dedicated to linear variable selection but numerous extensions have now emerged.
In structured sparsity, one attempts to estimate a function which, in a appropriate parameterization, is encoded by a sparse vector; the support (or set of non-zero elements) of this sparse vector is furthermore assumed to present a type of structure which is known a priori. A common approach to the problem is to penalize implicitly or explicitly the structure of the support of the estimated parameter vector.
In this talk, I will review the convex approaches to structured sparsity, some applications
and time permitting a generic optimal convex relaxation for a family of functions
that penalize simultaneously (a) the structure of the support through a general combinatorial set function, and (b) the Lp norm of the parameter vector for an arbitrary fixed p.