Stochastic Dual Coordinate Ascent and its Proximal Extension for Regularized Loss Minimization

Speaker  Tong Zhang

Affiliation  Rutgers University

Host  Dengyong Zhou

Duration  00:54:53

Date recorded  14 December 2012

Stochastic Gradient Descent (SGD) has become popular for solving large scale supervised machine learning optimization problems such as SVM, due to their strong theoretical guarantees. While the closely related Dual Coordinate Ascent (DCA) method has been implemented in various software packages, it has so far lacked good convergence analysis. We present a new analysis of Stochastic Dual Coordinate Ascent (SDCA) showing that this class of methods enjoy strong theoretical guarantees that are comparable or better than SGD. This analysis justifies the effectiveness of SDCA for practical applications.

Moreover, we introduce a proximal version of dual coordinate ascent method. We demonstrate how the derived algorithmic framework can be used for numerous regularized loss minimization problems, including L1 regularization and structured output SVM. The convergence rates we obtain match or improve state-of-the-art results.

Joint work with Shai Shalev-Shwartz

©2012 Microsoft Corporation. All rights reserved.
> Stochastic Dual Coordinate Ascent and its Proximal Extension for Regularized Loss Minimization