7a9a322cbe0d06a98667fdc5160dc6f8-AuthorFeedback.pdf

We provide extensive ablation stud-8 ies to verify the individual contribution of9 three complementary principles. R1: Additional cost of adopting multi-stage training and multiple discriminators Compared to the baselines,13 our DHA framework slightly increases the memory usage and computation at the training time. The according15 final performances are 28.8, 29.1, 29.5, and 32.0. Our framework aims tolearn domain-invariant51 representations that are robust on multiple latent target domains. As a result, the learned representations can well52 generalize on the unseen target domains by construction (please also refer to the Figure 1-(b)).