Speeding up Automatic Hyperparameter Optimization of Deep Neural Networks by Extrapolation of Learning Curves