Adversarial Discriminative Domain Adaptation

Adversarial Discriminative Domain Adaptation addresses unsupervised domain adaptation, where a model trained on a labeled source domain must generalize to an unlabeled target domain despite distribution shift. The authors first lay out a generalized framework for adversarial adaptation that subsumes several state-of-the-art approaches as special cases, distinguished by choices such as generative versus discriminative base models, tied versus untied weights, and the adversarial loss used. This unified view clarifies the trade-offs: GAN-based generative methods produce compelling images but are weaker on discriminative tasks and small shifts, whereas discriminative methods handle larger shifts but often constrain weights and forgo a GAN loss.

Guided by this framework, the paper proposes a previously unexplored combination, ADDA, that pairs discriminative modeling with untied weight sharing and a GAN-based loss. The method is both simpler and more effective than competing domain-adversarial techniques, and it surpasses prior state-of-the-art results on standard unsupervised adaptation benchmarks as well as a difficult cross-modality object classification task. By reconciling generative and discriminative adversarial adaptation, ADDA became a foundational reference for adversarial domain adaptation.