The article presents a novel approach to Unsupervised Domain Adaptation (UDA) called Invariant Consistency Learning (ICON). Traditional UDA methods often suffer from spurious correlations between domain-invariant and domain-specific features, which do not generalize well to the target domain. ICON addresses this by giving equal status to both the source and target domains. The prediction of an invariant classifier is simultaneously consistent with both the labels in the source domain and clusters in the target domain, thereby removing spurious correlation inconsistencies. The authors demonstrate through extensive experiments that ICON outperforms existing methods on various benchmarks.
Publication date: 22 Sep 2023
Project Page: https://github.com/yue-zhongqi/ICON
Paper: https://arxiv.org/pdf/2309.12742
