Robust Batch Relation Preserving for Lifelong Unsupervised Domain Adaptation

Lifelong unsupervised domain adaptation (lifelong UDA) is a challenging problem involving adaptation to unlabeled new domains, and alleviation of catastrophic forgetting of old domains. Existing methods impose an absolute relational distillation loss on features from the old model and the current model to alleviate forgetting, which hinders the model’s plasticity. Moreover, previous works in lifelong UDA naively employ pseudo-labeling mechanisms to learn from the unlabeled target domains, overlooking the noise problem in pseudo-labels. In this case, we propose Robust Batch Relation Preserving, allowing for more plasticity of the model when performing relational distillation and facilitating more robustness in pseudo-labeling mechanisms. Specifically, we consolidate previous knowledge by distilling batch-wise squished relation distributions, relaxing the constraints of rigid relation preserving. Meanwhile, we employ nuclear norm loss to promote diversity and improve certainty in intra-batch predictions to facilitate the robust pseudo-labeling of unlabeled target data. Extensive experiments on OfficeHome and DomainNet validate the effectiveness of our proposed method with a maximum improvement of 4.74% and 2.32% in the average adaptation performance and the overall accuracy after lifelong learning, respectively.