Heterogeneous domain adaptation by semantic distribution alignment network

Domain adaptation leverages the knowledge in the source domain to build a model that is capable of performing effectively in the target domain. Heterogeneous domain adaptation (HDA) is a special case of domain adaptation to deal with the situation that different types of features are used to represent the source and target domains. Most of the existing HDA works tackle this problem by using moment matching strategy such as Maximum Mean Discrepancy (MMD) to minimize the domain discrepancy. However, the heterogeneous domains can not always be completely characterized by lower-order statistics. Furthermore, the space complexity of calculating the higher-order statistics is high which makes it hard to calculate in practical applications. To tackle this problem, in this paper, we explore the advantages of utilizing Sliced Wasserstein Distance (SWD) to measure the distribution difference between the heterogeneous domains and propose a Semantic Distribution Alignment Network (SEDAN) for HDA. More specifically, the proposed SEDAN matches the difference of the marginal distribution between the two domains by using sliced Wasserstein distance and simultaneously minimizes the conditional distributions by using a novel Local Sliced Wasserstein Distance (LSWD) in a deep neural network. The LSWD captures the semantic information implicated in the unlabeled target domain by utilizing high confidence target samples, which guarantees the robustness of class-level domain matching. We provide theoretical analysis and also conduct complete experiments on two HDA tasks, the experimental results indicate that our SEDAN outperforms extant HDA baselines.