Ce vendredi, nous aurons un séminaire à 10h30 en salle 301.
Joseph Salmon (Télécom Paris) viendra nous faire un exposé intitulé:
Mind the (duality) gap: safer rules for the Lasso.
The Lasso is an l1-regularization technique for least-square, having the nice property of inducing sparse solutions. Recently, screening variables has been considered to speed up optimization algorithms solving a Lasso problem or its derivatives. We propose new versions of the so called safe rules for the Lasso. Based on duality gap considerations, our new rules create safe test regions whose diameters converge to zero for any solver outputting a converging sequence toward a Lasso solution. This helps screening out more variables for a wider range of tuning parameters. In addition to faster convergence to a solution, we prove that we correctly identify active sets (supports) in finite time. Though our framework can encompass any solver, we have implemented our strategies on a popular coordinate descent version. We achieved noticeable computing time reduction with respect to former safe rules.