Based on our findings, we propose a simple but effective method, Adversarial Training with Adaptive Step size (ATAS). ATAS learns an instance-wise adaptive step size that is inversely proportional to ...