Fetching the paper…

Small nonlinearities in activation functions create bad local minima in neural networks · Around