Learning Unforgotten Domain-Invariant Representations for Online Unsupervised Domain Adaptation

Cheng Feng, Chaoliang Zhong, Jie Wang, Ying Zhang, Jun Sun, Yasuto YokotaIJCAI 2022 · DOI

Introduction

Online Unsupervised Domain Adaptation (OUDA) presents a significant challenge in machine learning: adapting a model to a target domain using a continuous stream of small, unlabeled data batches without losing the knowledge acquired from the source domain. This paper addresses the issue of catastrophic forgetting in online scenarios, where standard offline Unsupervised Domain Adaptation (UDA) methods often fail due to parameter drift during sequential updates.

The Problem of Catastrophic Forgetting

In standard UDA, models typically have access to the entire target dataset at once. However, in real-world applications like robotics or autonomous driving, data arrives sequentially. When offline methods are applied to this online stream, the model tends to overfit to current small batches, destroying the robust features learned from the source. This creates a fundamental stability-plasticity trade-off: the model must be plastic enough to learn from new target data but stable enough to retain foundational source knowledge.

Methodology: Source Feature Distillation (SFD)

The authors introduce a new theoretical generalization bound based on Generator Discrepancy, which quantifies how much the feature generator changes between online steps. To minimize this discrepancy and prevent forgetting, they propose Source Feature Distillation (SFD).

Comparison between conventional UDA and the proposed SFD framework.

As illustrated in the framework comparison, SFD utilizes the original source-only model ($g^0$) as a "teacher." During online adaptation, the model is penalized via a distillation loss ($\mathcal{L}_{sfd}$) if its feature representations ($z$) drift too far from the teacher's representations. This "plug-and-play" module can be integrated into various UDA architectures like DANN, CDAN, and MCC with minimal computational overhead (approximately 0.15s per epoch).

Experimental Evaluation

The approach was evaluated on standard benchmarks including Office-31, Office-Home, and ImageCLEF-DA, simulating online environments with batches of 36 unlabeled samples.

Accuracy over online steps for DANN and CDAN with and without SFD.

As shown in the results, adding SFD significantly stabilizes accuracy across online steps. For instance, on the Office-Home dataset, SFD improved the final step accuracy of DANN from 49.5% to 57.2%. Qualitative analysis via T-SNE confirmed that SFD maintains a discriminative feature structure aligned with the source domain, reducing forgetting by approximately 70%.

Conclusion and Future Work

SFD provides a computationally efficient and theoretically grounded solution to catastrophic forgetting in OUDA. By acting as a "North Star" for feature alignment, it ensures that the model remains robust even as it adapts to new data. Future research may explore non-stationary source domains and adaptation in scenarios with extremely small batch sizes or single-sample updates.

Summary updated

BibTeX
@inproceedings{feng2022learning,
  author = {Cheng Feng and Chaoliang Zhong and Jie Wang and Ying Zhang and Jun Sun and Yasuto Yokota},
  title = {{Learning Unforgotten Domain-Invariant Representations for Online Unsupervised Domain Adaptation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/410},
}
Figure