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

Re-assign community