Unsupervised domain adaptation via reliable pseudolabeling based memory module and dynamic distance threshold learning

An important challenge in unsupervised domain adaptation (UDA) is how to capture domain-specific discriminative information in the target domain data. For the labels in the target domain are unavailable, some methods uncover the intrinsic discriminative information by clustering the target domain features in the embedding space. However, most of existing works only consider filtering reliable target domain samples by setting a fixed distance threshold, ignoring the potential intraclass and interclass relationships and cannot set appropriate distance thresholds for different transfer tasks and different classes. To this end, we propose a new pseudolabeling framework named the dynamic distance threshold learning framework (DDT) for UDA. In our framework, we construct distance memory module to encode intraclass and interclass distribution information, and select reliable target domain samples satisfying intraclass compactness and interclass independence via dynamic intraclass and interclass distance thresholds. Specifically, we define class centroid memory modules, intraclass and interclass distance memory modules in the embedding space. Then, we retrieve the intraclass compactness distance threshold and interclass independence distance threshold from the distance memory based on the class and classpair information of target domain sample. Finally, we use reliable target domain samples to update the class centroid and the distance memory module, so that the distance memory module is dynamically updated following the model training. Benefiting from DDT, the UDA model can be updated with reliable target domain samples for training, and the joint distribution can be further aligned at the category level. As a lightweight technique, DDT constructs class center and distance memory modules to transfer the intraclass and interclass distribution relationships from the source domain to the target domain, and sets dynamic intraclass and interclass distances for different classes in different transfer tasks. Without any additional training parameters, the framework can be easily plugged into various domain adaptation paradigms. Experiments on benchmark datasets show that our method can filter the noise in pseudolabels and confers a competitive advantage compared with existing state-of-the-art domain adaptation methods. & COPY; 2023 Elsevier B.V. All rights reserved.