Neural network activation functions

There is a whole cottage industry in showing neural networks are reasonably universal function approximators with various nonlinearities as activations, under various conditions. Usually we take it as a given that the particular activation function is not too important.

Sometimes, we might like to play with the precise form of the nonlinearities, even making the nonlinearities themselves directly learnable, because some function shapes might have better approximation properties with respect to various assumptions on the learning problems, in a sense which I will not attempt to make rigorous now, vague hand-waving arguments being the whole point of deep learning.

I think a part of this field has been subsumed into the stability-of-dynamical-systems setting? Or we do not care because something-something BatchNorm?

The current default activation function is ReLU, i.e. \(x\mapsto \max\{0,x\}\), which has many nice properties. However, it does lead to piecewise linear spline approximators. One could regard that as a plus (Unser 2019) but OTOH that makes it hard to solve differential equations.

Sometimes, then, we want something different. Other classic activations such as \(x\mapsto\tanh x\) have fallen from favour, supplanted by ReLU. However, differentiable activations are useful, especially if higher-order gradients of the solution will be important. Many virtues of differentiable activation functions are documented Implicit Neural Representations with Periodic Activation Functions. Sitzmann et al. (2020) argues for \(x\mapsto\sin x\) which has some handy properties. Ramachandran, Zoph, and Le (2017) advocate for Swish, \(x\mapsto \frac{x}{1+\exp -x}.\)

Other fun things, SELU, the “self-normalizing” SELU (scaled exponential linear unit) Klambauer et al. (2017).