Gradient descent provably escapes saddle points in the training of shallow ReLU networks

Open in new window