TIDOT: A Teacher Imitation Learning Approach for Domain Adaptation with Optimal Transport

Tuan Nguyen, Trung Le, Nhan Dam, Quan Hung Tran, Truyen Nguyen, Dinh Q. PhungIJCAI 2021 · OpenReview

Abstract

Unsupervised domain adaptation (UDA) aims to bridge the gap between a labeled source domain and an unlabeled target domain. We present Teacher Imitation Domain Adaptation with Optimal Transport (TIDOT), a novel framework that leverages imitation learning and optimal transport (OT) theory. TIDOT employs a cooperative teacher-student architecture where a student classifier learns to mimic an expert teacher's predictions on target samples by finding optimal matches in the source domain. By minimizing the OT distance between embedded distributions and predictive outputs, TIDOT effectively mitigates both covariate and label shifts. Empirical results demonstrate that TIDOT achieves state-of-the-art performance across multiple benchmark datasets, including significant improvements on challenging tasks like MNIST to SVHN.

Introduction

Domain adaptation is a critical challenge in machine learning, occurring when a model trained on a source distribution must generalize to a different target distribution. This discrepancy, often characterized by covariate shift (different data distributions) and label shift (different class distributions), causes standard models to perform poorly on real-world data. Traditional methods often focus on global feature alignment, which can fail to capture complex relationships between features and labels. TIDOT addresses this by framing UDA as an imitation learning problem, where the target-domain "student" learns from a source-domain "teacher" through a rigorous optimal transport mechanism.

Methodology

The TIDOT architecture consists of three core components: a shared feature generator ($G$), a source-domain expert teacher ($h_S$), and a target-domain student ($h_T$).

TIDOT Architecture and Clustering Logic

The Imitation Mechanism via Optimal Transport

The central innovation of TIDOT is the use of Optimal Transport to facilitate knowledge transfer. OT provides a principled way to find the most efficient matching between the source and target distributions in the latent space. Unlike methods that simply try to make distributions look identical, TIDOT uses OT to provide specific "demonstrations" for the student to follow.

Optimal Transport mapping between source and target distributions

As illustrated above, the mapping $K^*$ identifies source samples $\mathbf{x}_a$ that are most representative of target samples $\mathbf{x}_b$. The student model is then trained to minimize the discrepancy between its own predictions on target data and the teacher's predictions on the matched source data. This allows the student to inherit the teacher's expertise while adapting to the target domain's specific feature representations.

Refining Decision Boundaries

To ensure robustness, TIDOT incorporates a clustering assumption. By employing Virtual Adversarial Training (VAT) and entropy minimization, the model encourages decision boundaries to lie in low-density regions of the target feature space. This prevents the student from making ambiguous predictions near cluster edges, leading to more stable and confident classification. This consensus-seeking behavior also helps the model remain accurate even when class proportions differ between domains.

Experimental Evaluation

TIDOT was rigorously tested on standard UDA benchmarks, including digits (MNIST, USPS, SVHN), traffic signs (GTSRB), and complex object recognition datasets (Office-31, Office-Home, ImageCLEF-DA).

Key Results

  • State-of-the-Art Performance: TIDOT consistently outperformed existing methods. Notably, on the difficult MNIST $\to$ SVHN task, it achieved a 15.4% improvement over the previous best-performing model. On Office-31, it reached an average accuracy of 94.1%.
  • Student Superiority: A key finding was that the student classifier typically outperforms the teacher on target data. While the teacher provides the initial guidance, the student's direct exposure to target representations allows for superior adaptation.
  • Optimization Stability: The OT-based loss function exhibits smoother optimization compared to adversarial (GAN-based) approaches, which are often prone to erratic fluctuations.

Conclusion

TIDOT introduces a powerful paradigm for domain adaptation by combining the theoretical strengths of Optimal Transport with the practical benefits of imitation learning. By focusing on local sample matching rather than just global distribution alignment, TIDOT provides a more nuanced and effective way to transfer knowledge across domains. Future work will explore the computational efficiency of OT for larger datasets and the potential for "OT-inspired regularization" to mitigate overfitting in deep networks.

Summary updated

Abstract

Using the principle of imitation learning and the theory of optimal transport we propose in this paper a novel model for unsupervised domain adaptation named Teacher Imitation Domain Adaptation with Optimal Transport (TIDOT). Our model includes two cooperative agents: a teacher and a student. The former agent is trained to be an expert on labeled data in the source domain, whilst the latter one aims to work with unlabeled data in the target domain. More specifically, optimal transport is applied to quantify the total of the distance between embedded distributions of the source and target data in the joint space, and the distance between predictive distributions of both agents, thus by minimizing this quantity TIDOT could mitigate not only the data shift but also the label shift. Comprehensive empirical studies show that TIDOT outperforms existing state-of-the-art performance on benchmark datasets.
BibTeX
@inproceedings{nguyen2021tidot,
  author = {Tuan Nguyen and Trung Le and Nhan Dam and Quan Hung Tran and Truyen Nguyen and Dinh Q. Phung},
  title = {{TIDOT: A Teacher Imitation Learning Approach for Domain Adaptation with Optimal Transport}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=d5lsNMpjojD},
}
Figure