Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation

We consider unsupervised ___domain adaptation (UDA), where labeled data from a source ___domain (e.g., photographs) and unlabeled data from a target ___domain (e.g., sketches) are used to learn a classifier for the target ___domain. Conventional UDA methods (e.g., ___domain adversarial training) learn ___domain-invariant features to improve generalization to the target ___domain. In this paper, we show that contrastive pre-training, which learns features on unlabeled source and target data and then fine-tunes on labeled source data, is competitive with strong UDA methods. However, we find that contrastive pre-training does not learn ___domain-invariant features, diverging from conventional UDA intuitions. We show theoretically that contrastive pre-training can learn features that vary subtantially across domains but still generalize to the target ___domain, by disentangling ___domain and class information. Our results suggest that ___domain invariance is not necessary for UDA. We empirically validate our theory on benchmark vision datasets.