Stochastic Majorization-Minimization Optimization with First-Order Surrogate Functions - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2013

Stochastic Majorization-Minimization Optimization with First-Order Surrogate Functions

Résumé

Majorization-minimization algorithms consist of iteratively minimizing a majorizing surrogate of an objective function. Because of its simplicity and its wide applicability, this principle has been very popular in statistics and in signal processing. In this paper, we intend to make this principle scalable. We introduce and study a stochastic majorization-minimization scheme, which is able to deal with large-scale or possibly infinite data sets. When applied to convex optimization problems under suitable assumptions, we show that it achieves an expected convergence rate of O(1/\sqrt{n}) after n iterations, and O(1/n) for strongly convex functions. Equally important, our scheme almost surely converges to stationary points for a large class of non-convex problems. We derive from our framework several efficient algorithms. First, we propose a new stochastic proximal gradient method, which experimentally matches state-of-the-art solvers for large-scale l1-logistic regression. Second, we develop an online DC programming algorithm for non-convex sparse estimation. Finally, we demonstrate the effectiveness of our technique for solving large-scale structured matrix factorization problems.
Fichier principal
Vignette du fichier
online_surrogate.pdf (444.09 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-00835840 , version 1 (19-06-2013)
hal-00835840 , version 2 (10-09-2013)

Identifiants

Citer

Julien Mairal. Stochastic Majorization-Minimization Optimization with First-Order Surrogate Functions. 2013. ⟨hal-00835840v1⟩
673 Consultations
2605 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More