Automatic Gradient Descent: Deep Learning without Hyperparameters