Abstract
Machine learning models trained in one domain often perform poorly in others due to domain shift. Domain adaptation (DA) techniques address this by training transferable models from a label-rich source domain to a label-scarce target domain. However, most existing DA methods require target-domain data during training. This paper tackles the challenging problem of Zero-Shot Domain Adaptation (ZSDA), where target-domain data is completely unavailable during the training stage. We propose Conditional Coupled Generative Adversarial Networks (CoCoGAN), which extends the CoGAN architecture into a conditioning model. CoCoGAN captures the joint distribution of dual-domain samples across a relevant task (RT) and an irrelevant task (IRT). By training on source-domain RT samples and dual-domain IRT samples, CoCoGAN learns to "hallucinate" target-domain data for the RT. A novel representation alignment signal is introduced to ensure consistency across tasks. Extensive experiments demonstrate that CoCoGAN significantly outperforms state-of-the-art methods in image classification tasks.
Introduction: The Challenge of Zero-Shot Domain Adaptation
Domain shift is a fundamental challenge in machine learning, where a model's performance degrades when tested on data from a distribution different from its training set. While standard Domain Adaptation (DA) mitigates this using target-domain samples, Zero-Shot Domain Adaptation (ZSDA) addresses the more extreme scenario where no target data is available for the task at hand.
The proposed CoCoGAN framework, developed by researchers at Shenzhen University, leverages a secondary "Irrelevant Task" (IRT) for which dual-domain data is available. By learning the shared correlation between domains in the IRT, the model can apply this "bridge" to the "Relevant Task" (RT), effectively hallucinating the missing target data for training effective classifiers.
Methodology: The CoCoGAN Framework
CoCoGAN builds upon the Coupled GAN (CoGAN) architecture, designed to learn a joint distribution of multi-domain images without requiring corresponding image pairs.
Architecture and Weight Sharing
The framework consists of two parallel GANs: $GAN_s$ for the source domain and $GAN_t$ for the target domain. To force the networks to learn common semantic concepts (e.g., the shape of a digit), the generators and discriminators share weights in their high-level layers. Low-level layers remain unshared to capture domain-specific details like color or texture.
Task Conditioning and Representation Alignment
A binary conditioning variable $c$ allows the network to switch between the RT and the IRT. A critical innovation is the Representation Alignment supervisory signal. Task classifiers are integrated into the network, and the generators are trained to produce features that "fool" these classifiers. This ensures that the internal representations for both tasks are aligned within each domain, preventing the model from becoming biased toward the IRT and ensuring the domain transformation is universal.
Experimental Results
CoCoGAN was evaluated on datasets including MNIST, NIST, and Fashion-MNIST, with synthetic domain shifts such as colorization, edge detection, and intensity inversion.
- Performance Gains: CoCoGAN consistently outperformed the ZDDA baseline. In NIST letter classification using Fashion-MNIST as the IRT, it achieved a 34.0% improvement.
- Ablation Study: The representation alignment signal proved vital; its removal led to an average performance drop of 15.6% across all experiments.
- Qualitative Success: The model successfully generated synthetic target-domain samples that were visually indistinguishable in style from real target data while maintaining correct semantic labels from the source domain.
Conclusion and Future Work
CoCoGAN provides a robust solution for Zero-Shot Domain Adaptation by transferring domain knowledge across tasks. While highly effective, the method assumes a consistent domain transformation across tasks ($T_r = T_{ir}$). Future work may explore more complex domain shifts and higher-resolution natural scenes to further test the framework's scalability and robustness in diverse real-world applications.
Summary updated