University of Washington
In the last 10 years there has been active interest in finding sparse solutions to optimization problems. A popular approach is to use nonsmooth regularization (e.g., l_1-regularization) to induce solution sparsity, as exemplified by Basis Pursuit and lasso in statistical estimation. The regularized problem is large scale, sometimes dense, and nonsmooth. We discuss how block-coordinatewise descent methods can be applied to efficiently solve the regularized problem. Applications include variable selection in linear/logistic regression, sparse inverse covariance estimation. (This is joint work with Sylvain Sardy and Sangwoon Yun.)
University of Washington