Talk at SPOT (Multidisciplinary Optimization Seminar in Toulouse)

I’m giving a talk titled “Pushing the limits of Safe Screening” at SPOT 87 seminar on February 5 in Toulouse, France.

The slides for this 50 minute talk are available here.

The abstract of the presentation is the following:

Safe screening is a powerful tool to accelerate the convergence of sparse optimization solvers by performing early identification of zero coordinates while holding strong theoretical guarantees. Intially proposed for the Lasso, these techniques have later been extended to a larger range of problems including: non-negative and group Lasso, sorted-L1 penalized estimation (SLOPE), regularized logistic regression or even metric learning (via a regularized triplet loss minimization) and SVM. In this talk, we explore recent approaches that expand the applicability of safe screening in two main directions: 1) tackling a larger family of data-fidelity functions by relaxing global regularity hypotheses into local ones; 2) moving away from the sparsity contraint and leveraging the same formalism on box-constrained problems