Deep learning as optimal control problems: models and numerical methods.