Apple ML Research proposes anti-causal domain generalization method

The method uses unlabeled data from training environments to learn invariant predictors without requiring labeled data from multiple environments. It aims to improve robustness to distribution shifts in unseen domains.