A case for new neural network smoothness constraints