1: \begin{abstract}

2:   Stochastic Dual Coordinate Ascent is a popular method for solving

3:   regularized loss minimization for the case of convex losses.  We

4:   describe variants of SDCA that do not require explicit

5:   regularization and do not rely on duality. We prove linear

6:   convergence rates even if individual loss functions are non-convex,

7:   as long as the expected loss is strongly convex.

8: \end{abstract}

9: