Meta-Learning to Improve Pre-Training

Neural Information Processing Systems 

Choosing optimal PT hyperparameter values is challenging, and existing methods do not work well.