Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors Qixun Wang 1* Yifei Wang