Piecewise linear activations substantially shape the loss surfaces of neural networks

Open in new window