DA-NAS: Learning Transferable Architecture for Unsupervised Domain Adaptation

Recent studies have shown that unsupervised domain adaptation enables adapting models from a labeled source domain to a new unlabelled target domain. However, the existing methods only devoted to matching the feature distribution of the source and target domain, ignoring the discrepancy caused by improper architectures of feature extractors. While the network architectures are hand-crafted designed for image classification, we deem that extractors that are able to yield features with lower discrepancy between source and target domain are preferable for domain adaptation. In this work, we propose automatically searching neural networks for domain adaptation. To enhance feature transferability, we introduce a differentiable hierarchical search space, splitting our model into four stages with heterogeneous blocks. By further taking the architecture discrepancy among blocks in different stages into consideration, the searched architectures aim to facilitate a more effective transfer from the source domain to the target domain. The extensive experiments have validated our searched model on unsupervised domain adaptation.