Trainability and Data-dependent Initialization of Over-parameterized ReLU Neural Networks

Open in new window