Conditional Adversarial Domain Adaptation

Mingsheng Long, Zhangjie Cao, Jianmin Wang, Michael I. Jordan

Abstract

Adversarial learning has been embedded into deep networks to learn disentangled and transferable representations for domain adaptation. Existing adversarial domain adaptation methods may not effectively align different domains of multimodal distributions native in classification problems. In this paper, we present conditional adversarial domain adaptation, a principled framework that conditions the adversarial adaptation models on discriminative information conveyed in the classifier predictions. Conditional domain adversarial networks (CDANs) are designed with two novel conditioning strategies: multilinear conditioning that captures the cross-covariance between feature representations and classifier predictions to improve the discriminability, and entropy conditioning that controls the uncertainty of classifier predictions to guarantee the transferability. With theoretical guarantees and a few lines of codes, the approach has exceeded state-of-the-art results on five datasets.

1. Introduction

Deep learning models are notoriously sensitive to domain shift, where a model trained on a source domain fails to generalize to a target domain with a different distribution. Standard adversarial domain adaptation methods, such as Domain-Adversarial Neural Networks (DANN), attempt to learn domain-invariant features by training a discriminator to distinguish between source and target features. However, these methods typically align global distributions and may fail to account for the multimodal nature of classification tasks. In complex scenarios with many classes, global alignment can lead to "mode mismatch," where features from different classes are incorrectly aligned across domains—for instance, mapping source "dog" features to target "cat" features.

2. Conditional Domain Adversarial Networks (CDAN)

To address these challenges, the authors propose Conditional Domain Adversarial Networks (CDANs). The core innovation is conditioning the domain discriminator on the joint distribution of feature representations ($f$) and classifier predictions ($g$), ensuring that the alignment process is class-aware.

2.1 Multilinear Conditioning

Simple concatenation ($f \oplus g$) assumes independence between features and labels, which is often insufficient for capturing complex interactions. CDAN instead utilizes a multilinear map (outer product $f \otimes g$) to capture the cross-covariance between features and classes. To mitigate the "dimension explosion" associated with outer products, the authors implement a Randomized Multilinear Map using random feature mapping, maintaining computational efficiency without sacrificing performance.

Architecture of CDAN with multilinear and randomized multilinear conditioning strategies.

2.2 Entropy Conditioning

Not all target samples are equally informative for adaptation. CDAN introduces entropy conditioning (CDAN+E), which weights samples based on the uncertainty of the classifier's predictions. By prioritizing samples with low-entropy (high-certainty) predictions, the model effectively filters out noisy or ambiguous target data that could otherwise degrade the learned feature representations.

3. Experimental Results

The CDAN framework was evaluated on five major benchmarks: Office-31, ImageCLEF-DA, Office-Home, Digits, and VisDA-2017.

Performance comparison of CDAN and CDAN+E against state-of-the-art methods on ImageCLEF-DA, Office-Home, Digits, and VisDA-2017 datasets.

Key findings include:

  • Office-Home: On this challenging 65-class dataset, CDAN+E achieved an average accuracy of 65.8%, a significant improvement over the previous state-of-the-art (JAN at 58.3%).
  • VisDA-2017: CDAN+E reached 70.0% accuracy, outperforming complex generative models like GTA (69.5%) despite being a simpler discriminative approach.
  • Office-31: The model achieved 87.7% average accuracy using a ResNet-50 backbone, surpassing the previous best of 86.5%.

Qualitative analysis using t-SNE visualizations confirmed that CDAN produces tighter, class-distinct clusters that are more precisely aligned between domains compared to standard adversarial methods.

4. Conclusion

CDAN provides a robust and theoretically grounded approach to domain adaptation by leveraging classifier predictions to guide adversarial alignment. By combining multilinear conditioning for class-awareness and entropy weighting for uncertainty control, it sets new performance standards across diverse and challenging datasets. The simplicity and effectiveness of the approach make it a powerful tool for improving the generalization of deep learning models in real-world applications.

Summary updated

Abstract

Adversarial learning has been embedded into deep networks to learn disentangled and transferable representations for domain adaptation. Existing adversarial domain adaptation methods may not effectively align different domains of multimodal distributions native in classification problems. In this paper, we present conditional adversarial domain adaptation, a principled framework that conditions the adversarial adaptation models on discriminative information conveyed in the classifier predictions. Conditional domain adversarial networks (CDANs) are designed with two novel conditioning strategies: multilinear conditioning that captures the cross-covariance between feature representations and classifier predictions to improve the discriminability, and entropy conditioning that controls the uncertainty of classifier predictions to guarantee the transferability. With theoretical guarantees and a few lines of codes, the approach has exceeded state-of-the-art results on five datasets.
BibTeX
@inproceedings{long2018conditional,
  author = {Mingsheng Long and Zhangjie Cao and Jianmin Wang and Michael I. Jordan},
  title = {{Conditional Adversarial Domain Adaptation}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2018},
  url = {https://openreview.net/forum?id=H1WdytWdbH},
}
Figure