SAGA: A Fast Incremental Gradient Method With Support for Non-Strongly Convex Composite Objectives (original) (raw)

View PDF

Abstract:In this work we introduce a new optimisation method called SAGA in the spirit of SAG, SDCA, MISO and SVRG, a set of recently proposed incremental gradient algorithms with fast linear convergence rates. SAGA improves on the theory behind SAG and SVRG, with better theoretical convergence rates, and has support for composite objectives where a proximal operator is used on the regulariser. Unlike SDCA, SAGA supports non-strongly convex problems directly, and is adaptive to any inherent strong convexity of the problem. We give experimental results showing the effectiveness of our method.

Submission history

From: Francis Bach [view email] [via CCSD proxy]
[v1] Tue, 1 Jul 2014 11:47:56 UTC (715 KB)
[v2] Tue, 22 Jul 2014 06:57:50 UTC (716 KB)
[v3] Tue, 16 Dec 2014 08:44:27 UTC (725 KB)