Unsupervised Subdomain Adaptation Robust to Noisy Pseudo Labels Using Symmetric Loss

Abstract

Abstract Alignment of subdomain distribution plays an important role in preventing negative transfer in domain adaptation. Due to the lack of labeled data in the target domain, the current mainstream methods prefer using pseudo labels to align the features of corresponding categories between the source and the target domains. However, the noises present in pseudo labels affect the effectiveness of subdomain alignment. In the community of domain adaptation, samples with higher confidences are believed to be more reliable when generating pseudo labels. However, we found that this conclusion does not hold for hard samples. To address this issue, we propose a pseudo label screening mechanism which considers the trade-off of quantity and quality. The qualified target samples participate in the subdomain adaptation, while the unqualified samples are randomly assigned with labels. Thus, the asymmetric noises of hard samples are converted into symmetric noises. Symmetric loss is proved to be robust to symmetric noises. Inspired by this observation, we propose a symmetric subdomain adaptation loss (SSAL) and construct a robust subdomain adaptation network (RSAN) based on SSAL accordingly. Leveraging the random label assignments of hard samples and SSAL, we reconstruct the relation between sample’s confidence and the probability being correctly classified. The effectiveness of our method has been validated on public benchmarks. Compared with the SOTA method, our RSAN obtains an improvement of 2.7% in terms of average accuracy on the challenging VisDA-2017 transfer task.