Unknown-Aware Domain Adversarial Learning for Open-Set Domain Adaptation
JoonHo Jang,Byeonghu Na,Dong Hyeok Shin,Mingi Ji,Kyungwoo Song,Il-chul Moon
Open-Set Domain Adaptation (OSDA) assumes that a target domain contains unknown classes, which are not discovered in a source domain. Existing domain adversarial learning methods are not suitable for OSDA because distribution matching with $ extit{unknown}$ classes leads to negative transfer. Previous OSDA methods have focused on matching the source and the target distribution by only utilizing $ extit{known}$ classes. However, this $ extit{known}$-only matching may fail to learn the target-$ extit{unknown}$ feature space. Therefore, we propose Unknown-Aware Domain Adversarial Learning (UADAL), which $ extit{aligns}$ the source and the target-$ extit{known}$ distribution while simultaneously $ extit{segregating}$ the target-$ extit{unknown}$ distribution in the feature alignment procedure. We provide theoretical analyses on the optimized state of the proposed $ extit{unknown-aware}$ feature alignment, so we can guarantee both $ extit{alignment}$ and $ extit{segregation}$ theoretically. Empirically, we evaluate UADAL on the benchmark datasets, which shows that UADAL outperforms other methods with better feature alignments by reporting state-of-the-art performances.
