This research tackles challenges in unsupervised domain adaptation (UDA), where the goal is to perform well in a distribution shifted domain without labels. The authors develop an importance weighted group accuracy estimator to improve model calibration and selection under these conditions. The study highlights the significance of group accuracy estimation in addressing UDA challenges and emphasizes its potential in improving the transferability of accuracy. The paper also discusses the difficulties in enhancing calibration performance in UDA due to distribution shifts and the challenges in model selection due to the scarcity of labeled target domain data.
Publication date: 17 Oct 2023
Project Page: https://arxiv.org/abs/2310.10611
Paper: https://arxiv.org/pdf/2310.10611
