Robust Domain Adaptation: Representations, Weights and Inductive Bias