Learning Activation Functions for Sparse Neural Networks