Paper accepted at ECCV2020
It is our pleasure to announce that one of our papers has been accepted as a poster to ECCV2020, which will be held online. It is a collaborative work with Prof. Masaaki Iiyama, Kyoto University.
Ryuhei Takahashi, Atsushi Hashimoto, Motoharu Sonogashira, Masaaki Iiyama, “Partially-Shared Variational Auto-encoders for Unsupervised Domain Adaptation with Target Shift” [preprint]
What is Unsupervised Domain Adaptation (UDA)?: UDA is one of the recent hottest topics in the machine learning community. It is a problem to make a model adapt to an environment where ground truth labels have not been provided.
It assumes two environments: source domain and target domain. In the source domain, you have plenty of ground truth labels. In the target domain, you have no ground truth labels but have many unlabeled observations. The performance is evaluated on the accuracy in the target domain.
Typical usages of our new method: When you would like to train a model with CG data for regression but apply it to the real-world, this method will largely help you realizing the application. Regression is a problem to estimate continuous values (e.g. body/hand/head pose estimation, gaze estimation, face landmark estimation, emotion strength estimation, age estimation, and any other continual parameter estimation from images).
Note that the method is also applicable to categorization problems. In that case, the method will get the advantage when you can expect a large difference in the label distributions of two domains.
Advantage: We believe that this method is the only method applicable to regression with UDA under unknown label distribution in the target domain. It has also another advantage in hyper-parameter tuning. Many UDA methods have a critical problem in hyper-parameter tuning. Namely, it must tune sensitive parameters without accessing ground truth labels. This is normally ill-posed and a hard task. In contrast, the proposed method output synthesized target domain data and the quality of the synthesis directly contributes to the performance. Hence, you can intuitively adjust hyperparameters by comparing the synthesized target domain samples and real target domain observations.
limitation: This method uses CycleGAN to convert source domain images to target domain ones. CycleGAN’s scope is not only images but our method does not work when failing to convert domains by any CycleGAN method.
Technical Details: The above paper has pointed out a critical problem of the traditional methods; they assume that the unseen label in the target domain distributes samely as the seen labels in the source domain. This is a dangerous assumption in a real problem and it actually degrades the UDA result.
For achieving UDA in a practical situation, we developed a new network architecture, “Partially-shared Variational Auto-Encoders: PS-VAEs.” This architecture avoids matching the distribution shape of the features, which almost all recent UDA methods implicitly try to do, but synthesize paired observation in the two domains with an identical label. Using such data, metric learning enables us to align features without relying on the matching of distribution shape. For more detail, please see the paper or visit our presentation in ECCV2020 or MIRU2020(a Japanese domestic meeting)!
