Piecewise linear activations substantially shape the loss surfaces of neural networks.

Fengxiang He, Bohan Wang, Dacheng Tao

Browse the full ICLR paper archive.