Deep Patel

We provide the first convergence guarantees and overparameterization bounds for alternating gradient methods in a special class of nonconvex-nonconcave zero-sum games where both the players are shallow neural networks.