1: \begin{abstract}

2: %%% JE suggested shorter abstract

3: We present a new, stochastic variant of the projective splitting (PS) family

4: of algorithms for monotone inclusion problems.  It can solve min-max and

5: noncooperative game formulations arising in applications such as

6: robust ML without the convergence issues associated with gradient

7: descent-ascent, the current \emph{de facto} standard approach in such

8: situations.  Our proposal is the first version of PS able to use stochastic

9: (as opposed to deterministic) gradient oracles. It is also the first stochastic

10: method that can solve min-max games while easily handling multiple constraints

11: and nonsmooth regularizers via projection and proximal operators.  We close with numerical experiments on a

12: distributionally robust sparse logistic regression problem.

13: %%% Old abstract

14: % Optimization has always played a key role in machine learning (ML). While in the past, the main focus was minimizing a single loss, much attention has now turned to more complicated min-max and noncooperative game formulations arising in applications such as GANs and robust ML.  Developing convergent methods for these games is notoriously difficult and the de facto standard method, gradient descent-ascent, is plagued by convergence issues. The recently-developed projective splitting method (PS) does not suffer from these convergence issues on convex games and can also handle multiple nonsmooth regularizers and constraints. However, PS requires a deterministic gradient oracle, making it impractical for most ML applications. In this work, we rectify this deficit and develop a stochastic PS method.  This is  Experiments are conducted on a distributionally-robust sparse logistic regression problem.

15:   % On the numerical side, the method significantly outperforms deterministic splitting methods on a distributionally-robust sparse logistic regression problem.

16: \end{abstract}

17: