Discriminative active learning for domain adaptation

Domain Adaptation aiming to learn a transferable feature between different but related domains has been well investigated and shown excellent empirical performances. Previous works mainly focused on matching the marginal distributions using adversarial training methods while assuming conditional relations source target domain remained unchanged, i.e., ignoring shift problem. However, recent have that such problem exists can hinder adaptation process. To address this issue, we leverage labeled data from domain, collecting be quite expensive time-consuming. end, introduce discriminative active learning approach for reduce efforts of annotation. Specifically, propose three-stage neural networks: invariant space (first stage), uncertainty diversity criteria their trade-off query strategy (second stage) re-training with queried labels (third stage). Empirical comparisons existing four benchmark datasets demonstrate effectiveness proposed approach. Furthermore, by comparing strategies, could benefits our method.