Learning from Teaching Regularization: Generalizable Correlations Should be Easy to Imitate

Neural Information Processing Systems 

Generalization remains a central challenge in machine learning.