Currently only TanH, Linear, and Gaussian radial basis functions are available for activation functions. RELUs (Rectified linear units) have become increasingly common, as they can be used to prune a neural network as well. The activation function is f(x) = max(0,x). Thus, if a layer uses this function and it outputs all 0's, then it could be pruned, since it won't affect the output(s).
