Abstract
Unsupervised domain adaptation (UDA) reduces the reliance on data annotation by adapting knowledge from a source to a target domain. Source-free domain adaptation (SFDA) extends this by adapting a pre-trained source model to an unlabeled target domain without accessing the source data, addressing privacy and efficiency concerns. However, most existing SFDA methods focus on the transductive setting, where the target training set is also the testing set. This paper addresses the more realistic inductive setting, where training and testing sets are mutually exclusive. We propose Dual Moving Average Pseudo-Labeling (DMAPL), a semi-supervised fine-tuning method that splits the unlabeled target training set into confident and less-confident subsets. By employing a soft-label moving-average updating strategy and a moving-average prototypical classifier, DMAPL gradually adapts the source model toward the target domain, achieving state-of-the-art performance.
Introduction
In modern machine learning, the ability to transfer knowledge across domains is critical. While Unsupervised Domain Adaptation (UDA) has traditionally relied on simultaneous access to source and target data, the emerging field of Source-Free Domain Adaptation (SFDA) operates under the constraint that only a pre-trained source model is available. This is particularly relevant for industries where data privacy (e.g., medical records) or data volume prevents the sharing of original source datasets.
A significant limitation in current SFDA research is the prevalence of transductive evaluation. In transductive settings, the model is evaluated on the same unlabeled data it used for adaptation. This paper argues for a shift toward the inductive setting, where the model must generalize to unseen target data. Without this distinction, models risk overfitting to the specific noise of the training samples, failing to provide robust performance in real-world deployment.
The DMAPL Framework
To address the challenges of inductive adaptation, we introduce Dual Moving Average Pseudo-Labeling (DMAPL). The framework utilizes a "divide and conquer" strategy to manage the uncertainty inherent in unlabeled target data through two primary mechanisms.
1. Confidence-Based Data Partitioning
DMAPL begins by evaluating the unlabeled target training set using the pre-trained source model. The data is partitioned into two distinct subsets based on prediction confidence scores:
- Confident Subset ($\mathcal{L}$): These instances are treated as "anchors." By assigning fixed pseudo-labels to these high-confidence samples, the model preserves the foundational knowledge of the source domain and prevents catastrophic forgetting.
- Less-Confident Subset ($\mathcal{U}$): These instances are used for semi-supervised adaptation. Because the source model is uncertain about these samples, they require a more nuanced approach than simple hard-labeling.
2. Dual Moving Average (DMA) Mechanism
To mitigate the noise in the less-confident subset, DMAPL employs two levels of temporal smoothing:
- Moving-Average Prototypical Classifier: Rather than recalculating class centers (prototypes) from scratch for every batch, the model maintains a moving average of class centers ($\mu_c$). This ensures that the geometry of the feature space evolves smoothly as the model adapts.
- Soft-Label Updating: To prevent the model from oscillating between different class assignments, we maintain a moving average of predicted labels ($q_i$). This soft-labeling approach provides a more stable training signal than "one-hot" hard labels, allowing the model to gradually drift toward the target domain distribution.
Experimental Evaluation
We evaluated DMAPL on two major benchmarks designed to test inductive performance:
- DomainNet: A large-scale dataset featuring 345 categories across domains such as clipart, painting, and real photos. We conducted 12 tasks across 4 domains.
- VisDA2017Split: A synthetic-to-real challenge. To ensure a proper inductive evaluation, we created a standard 8:2 train/test split for the target domain, as the original dataset lacks a standard inductive partition.
Results and Insights
DMAPL achieved state-of-the-art results across both benchmarks:
- DomainNet: DMAPL reached an average accuracy of 55.4%, significantly outperforming previous source-free methods like SSFT-SSD (50.4%) and even surpassing standard UDA methods like CDAN (49.9%), which benefit from source data access.
- VisDA2017Split: The method achieved 83.0% Macro accuracy, a substantial improvement over the source-only baseline of 48.9%.
Ablation studies confirmed that both the data splitting and the soft-label updating are vital; removing either component resulted in a performance drop of 3% to 7%. Furthermore, hyperparameter analysis indicated that slower updates (higher $\alpha$ and $\beta$ values) lead to better generalization, suggesting that stability is more critical than rapid convergence in the inductive setting.
Conclusion
DMAPL demonstrates that source-free inductive domain adaptation is not only possible but can exceed the performance of traditional UDA methods. By focusing on stability through dual moving averages and anchoring the model with high-confidence samples, DMAPL provides a robust framework for real-world applications where source data is unavailable and generalization to unseen data is paramount. Future work will focus on narrowing the remaining gap to "Oracle" performance (75.7% on DomainNet) and automating hyperparameter tuning for diverse domains.
Summary updated