unsupervised domain adaptation (UDA) methods for learning domain invariant
representations have achieved remarkable progress. However, most of the studies
were based on direct adaptation from the source domain to the target domain and
have suffered from large domain discrepancies. In t