event
Thursday 07 Jun 2012: Sparse prediction, matrix completion and first-order optimization
Dr. Andreas Argyriou - Toyota technological Institute at Chicago
Harrison 170 15:00-16:00
We derive a novel norm, the k-support norm, which corresponds
to the tightest convex relaxation of sparsity combined with
an L2 penalty. We show that this new norm provides a tighter
relaxation to support recovery than the elastic net and
suggest using it as a replacement for the Lasso or the elastic
net in sparse prediction problems. We also present a first-order
algorithm for learning with this norm.
In addition, we propose a general-purpose algorithm for convex
optimization problems involving Lipschitz and nonsmooth functions.
This method requires only first-order computations, like gradients,
and hence can scale well to large data sets. We demonstrate its
efficiency and wide applicability on problems like compressed sensing,
matrix completion and robust PCA.
