Meta-Learning Loss Functions for Deep Neural Networks