Partial domain adaptation based on shared class oriented adversarial network

摘要

Most existing domain adaptation methods assume that the label space of the source domain is the same as the label space of the target domain. However, this assumption is generally untenable due to the differences between the two domains. Therefore, a novel domain adaptation paradigm called Partial Domain Adaptation (PDA), which only assumes that the source label space is large enough to subsume the target label space has been proposed recently to relax such strict assumption. Previous partial domain adaptation methods mainly utilize weighting mechanisms to alleviate negative transfer caused by outlier classes samples. Though these methods have achieved high performance in PDA tasks, all the heterogeneous data is retained during the whole training process, which still contributes to negative transfer. In this work, we propose a shared class oriented adversarial network (SCOAN) for partial domain adaptation. Outlier samples are excluded from training process via weighting strategy to entirely circumvent negative transfer and positive transfer is performed by combining adversarial network and Maximum Mean Discrepancy (MMD) to bridge domain gap. Multi-classifier module is proposed to further improve the generalization ability of the network. Extensive experiments show that SCOAN achieves state-of-the-art results on several benchmark partial domain adaptation datasets.