Abstract
Generalizing deep neural networks to new target domains is critical to their real-world utility. In practice, it may be feasible to get some target data labeled, but to be cost-effective it is desirable to select a maximally-informative subset via active learning (AL). We study the problem of AL under a domain shift, called Active Domain Adaptation (Active DA). We demonstrate how existing AL approaches based solely on model uncertainty or diversity sampling are less effective for Active DA. We propose Clustering Uncertainty-weighted Embeddings (CLUE), a novel label acquisition strategy for Active DA that performs uncertainty-weighted clustering to identify target instances for labeling that are both uncertain under the model and diverse in feature space. CLUE consistently outperforms competing label acquisition strategies for Active DA and AL across learning settings on 6 diverse domain shifts for image classification.
Introduction: The Challenge of Domain Shift
Deep learning models often struggle when deployed in environments that differ from their training data, a phenomenon known as domain shift. While labeling a small portion of the target domain data can help, the goal of Active Domain Adaptation (Active DA) is to select the most informative samples within a limited budget.
Traditional Active Learning (AL) strategies typically fall into two categories: uncertainty-based and diversity-based sampling. However, these methods often falter under domain shift. Uncertainty-based sampling may select miscalibrated outliers or redundant instances, while diversity-based sampling might pick well-aligned "easy" instances that provide little new information.
Methodology: Clustering Uncertainty-weighted Embeddings (CLUE)
To address these limitations, we propose CLUE. The core idea is to identify target instances that are both informative (high uncertainty) and representative of the target distribution (high diversity).
The CLUE pipeline consists of three main stages:
- Target Uncertainty Weighting: The model calculates predictive entropy for all unlabeled target data to quantify uncertainty.
- Uncertainty-weighted Clustering: We perform clustering in the embedding space, where the density is weighted by the calculated uncertainty. This ensures that cluster centers are biased toward regions where the model is most confused.
- Representative Selection: The instances closest to the resulting cluster centroids are selected for human labeling.
This approach ensures that the selected budget is spread across different regions of the feature space (diversity) while focusing on areas of high model error (uncertainty).
Experimental Results
We evaluated CLUE across six diverse domain shifts, including benchmarks like DomainNet, DIGITS, and Office. CLUE consistently outperformed baselines such as entropy sampling, coreset, and AADA.
Key findings include:
- Robustness: CLUE provides significant gains even on difficult shifts like Clipart → Quickdraw, improving accuracy by up to 9% over previous state-of-the-art methods.
- Efficiency: By clustering in the lower-dimensional penultimate layer embedding space rather than high-dimensional gradient spaces (as in methods like BADGE), CLUE remains computationally efficient and avoids the "curse of dimensionality."
- Flexibility: CLUE is "back-end agnostic," meaning it can be integrated with various training strategies, from simple fine-tuning to complex domain-adversarial methods.
Conclusion
CLUE offers a principled and effective approach to Active Domain Adaptation by balancing uncertainty and diversity through weighted clustering. Its computational efficiency and consistent performance across diverse datasets make it a practical choice for real-world applications where labeling resources are scarce. Future work may involve extending this framework to more complex tasks like object detection and exploring more robust uncertainty measures.
Summary updated