Reviews: Generalizing to Unseen Domains via Adversarial Data Augmentation

The paper attacks a novel problem: one-shot domain generalization where given samples from a single domain one requires robustness to unknown domain covariate shift. The problem is extremely hard and the related works claim that no other work exists attacking the same problem. The paper incrementally builds up on a recently proposed method to defend against adversarial attacks [33]. In fact, the work uses the formulation of [33] and repurposes the procedure for one-shot domain generalization. The original additions are 3-fold: 1) the closeness constraint is changed from pixel-space to feature-space 2) an ensemble of models are trained with different neighborhood thresholds 3) a new theoretical motivation is provided.