 This paper proposes a new operator splitting algorithm that can directly solve a convex relaxation of the worst case performance of deep neural networks. This algorithm is able to scale to very large problems and is suitable for parallelization on GPUs. It has been tested in image classification and reinforcement learning tasks, as well as in reachability analysis of neural network dynamical systems. This article was authored by Xiaorou Chen, Eric Wang, J. Zico Koulter, and others.