Noisy-labeled Domain Adaptation under Generalized Label Shift

Abstract

The recent high performance of many deep learning models heavily relies on the massive amount of labeled data for training. Still, the correct annotation may be expensive or even impossible in practice. Unsupervised domain adaptation aims to train a robust classifier when we have access to unlabeled samples from a target domain and labeled samples from a source domain, which has been intensively studied in the literature. However, it is often relatively easy to obtain additional noisy labels from the target domain by, e.g., heuristic labeling, crowdsourcing, and the prediction from a preliminary model. This paper considers a novel problem setting where the target samples are equipped with noisy labels and proposes a method to incorporate the noisy target labels in the generalized label shift framework. We evaluate its performance via thorough experiments on several benchmark datasets and show that it can help transfer knowledge from the source to the target domains.