ADA (Anchor Data Augmentation) is evaluated and compared to prior approaches on tasks involving out-of-distribution robustness.
Four out-of-distribution datasets are used, including RCFashionMNIST, Communities and Crime, SkillCraft1 Master Table, and Drug-target Interactions.
Results show that ADA is competitive with C-Mixup and other data augmentation strategies, and significantly outperforms them on the SkillCraft dataset.
ADA reduces the error by around 15% compared to the ERM solution on the SkillCraft data.