LGFeb 4, 2016
SDCA without Duality, Regularization, and Individual Convexity
arXiv:1602.01582v2108 citations
Originality Incremental advance
AI Analysis
This work is incremental, improving SDCA's applicability to more general optimization problems in machine learning.
The authors tackled the problem of extending Stochastic Dual Coordinate Ascent (SDCA) to settings without explicit regularization, duality, or individual convexity, and proved linear convergence rates for non-convex individual loss functions under expected strong convexity.
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.