SDCA without Duality, Regularization, and Individual Convexity
Machine Learning
2016-05-24 v2
Abstract
Stochastic Dual Coordinate Ascent is a popular method for solving regularized loss minimization for the case of convex losses. We describe variants of SDCA that do not require explicit regularization and do not rely on duality. We prove linear convergence rates even if individual loss functions are non-convex, as long as the expected loss is strongly convex.
Keywords
Cite
@article{arxiv.1602.01582,
title = {SDCA without Duality, Regularization, and Individual Convexity},
author = {Shai Shalev-Shwartz},
journal= {arXiv preprint arXiv:1602.01582},
year = {2016}
}
Comments
ICML 2016