1: \begin{abstract}

2:   We present a new Frank-Wolfe (FW) type algorithm that

3:   is applicable to minimization problems with a nonsmooth convex objective.

4:   We provide convergence bounds and show that the scheme yields so-called {\em coreset} results for various Machine

5:   Learning

6:   problems including

7:   1-median, Balanced Development, Sparse PCA,

8:   Graph Cuts,

9:   and the $\ell_1$-norm-regularized Support Vector Machine (SVM) among others.

10:   This means that the algorithm provides approximate solutions to these problems

11:   in time complexity bounds that are {\em not dependent} on the size of the

12:   input problem. 

13:   Our framework, motivated by a growing body of work on {\em sublinear} algorithms for various data analysis problems, 

14:   is entirely deterministic and makes no use of smoothing or proximal operators. 

15:   Apart from these theoretical results, we 

16:   show experimentally

17:   that the algorithm is very practical and in some cases also offers significant computational advantages on large problem instances. 

18:   We provide an open source implementation that can be adapted for other problems that fit the 

19:   overall structure.

20: \end{abstract}

21: