Choosing activation functions for feedforward neural networks