Comment by jing

In practice the functions just need to be piecewise differentiable. The RELU is the canonical example for deep learning. At kinks a subderivative is used.