Source-linked AI summary
Attracting and Dispersing: A Simple Approach for Source-free Domain Adaptation
Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, Joost van de Weijer
TL;DR
Source-free domain adaptation must adapt models to an unlabeled target domain without source data, while existing approaches can be noisy, complex, or discard dissimilar-sample information. AaD treats SFDA as unsupervised clustering, optimizing an upper bound that attracts predictions of neighboring features and disperses predictions of dissimilar features; it achieves state-of-the-art results and extends to open-set and partial-set settings.
Problem
SFDA adapts models without source data, but existing methods can suffer from noisy pseudo-labels, ignore dissimilar samples, require complex processing, or store historical models.
Method
AaD treats SFDA as unsupervised clustering and optimizes an upper bound with discriminability and diversity terms for similar and dissimilar target features.
Results
AaD achieves state-of-the-art performance on several benchmarks, improving VisDA by 2.1% to 88.0% and extending to source-free open-set and partial-set adaptation.
Takeaways & Limitations
AaD provides a simple SFDA baseline with reported effectiveness across closed-set benchmarks and source-free open-set and partial-set settings.
Takeaways & Limitations
The method assumes a source-pretrained model and target data, with the paper's formal setting using the same C classes in source and target domains.
Abstract
from arXiv · showhide
We propose a simple but effective source-free domain adaptation (SFDA) method. Treating SFDA as an unsupervised clustering problem and following the intuition that local neighbors in feature space should have more similar predictions than other features, we propose to optimize an objective of prediction consistency. This objective encourages local neighborhood features in feature space to have similar predictions while features farther away in feature space have dissimilar predictions, leading to efficient feature clustering and cluster assignment simultaneously. For efficient training, we seek to optimize an upper-bound of the objective resulting in two simple terms. Furthermore, we relate popular existing methods in domain adaptation, source-free domain adaptation and contrastive learning via the perspective of discriminability and diversity. The experimental results prove the superiority of our method, and our method can be adopted as a simple but strong baseline for future research in SFDA. Our method can be also adapted to source-free open-set and partial-set DA which further shows the generalization ability of our method. Code is available in https://github.com/Albert0147/AaD_SFDA.
1 Introduction
Domain shift limits model generalization when target data differ from training distributions, motivating source-free adaptation methods that operate without source data. AaD addresses this setting with a simple clustering objective, reports state-of-the-art results, and extends to open-set and partial-set adaptation.
- Motivation: Domain shift causes substantial performance degradation when models trained on one distribution are applied directly to a different domain.Differences may involve backgrounds, styles, or camera parameters.
- Motivation: SFDA adapts models to an unlabeled target domain without source data, partly addressing data privacy and intellectual property concerns.
- Motivation: Existing SFDA methods may suffer from noisy pseudo-labels, ignore dissimilar-sample information, require complex processing, or store historical models.
- Contribution: AaD encourages similar target features to have similar predictions and dissimilar features to have dispersed predictions through an upper-bounded objective with discriminability and diversity terms.The method treats SFDA as unsupervised clustering and simultaneously targets feature clustering and cluster assignment.
- Results: 2.1% to 88.0%: AaD improves state-of-the-art performance on VisDA and also extends to source-free open-set and partial-set domain adaptation.
- Contribution: AaD relates domain adaptation, SFDA, and contrastive learning methods through the perspectives of discriminability and diversity.The authors present this connection as useful for understanding existing methods and future improvement.
2 Related Work
Prior work addresses domain adaptation through distribution alignment, adversarial learning, clustering, or source-free alternatives that synthesize data, use auxiliary models, or optimize feature assignments. AaD is positioned among these approaches as a simple method using target-feature memory and nearest-neighbor retrieval.
- Domain Adaptation: Traditional domain adaptation methods align feature distributions, use adversarial training, or exploit prediction diversity across classifiers.
- Source-free Domain Adaptation: Source-free methods synthesize training samples, freeze source classifiers, use pseudo-labeling, generate target-style images, or learn additional target-specific classifiers.
- Source-free Domain Adaptation: AaD requires a source-pretrained model and target data while adapting without access to source-domain examples.
- AaD Procedure: AaD's adaptation algorithm stores target features and predictions in a memory bank, retrieves K-nearest neighbors, and updates the model using its objective.
- Deep Clustering and Contrastive Learning: Deep clustering methods differ in whether feature representations and cluster assignments are learned simultaneously or alternately.Examples include alternating similarity and assignment updates, mutual-information objectives, and augmentation-based methods.
3 Method
AaD formulates source-free domain adaptation as unsupervised clustering by aligning predictions of nearby features and separating predictions of distant features. An approximated upper-bound yields a simple objective whose terms promote discriminability and diversity, while relating AaD to existing methods.
- Attracting and Dispersing: AaD clusters target features by attracting predictions of nearby features and dispersing predictions of farther-away features.The method uses the source-pretrained model's semantic structure and operates on predictions rather than directly on features.
- Neighborhood Construction: For each feature, the close-neighbor set contains its K-nearest features, while the background set contains features outside that neighborhood.Nearest neighbors are retrieved using cosine similarity from memory banks storing target features and predictions.
- Objective: The AaD negative log-likelihood jointly encourages similar predictions for close neighbors and less similar predictions for background features.Using classifier weights as class prototypes also supports feature clustering and cluster assignment simultaneously.
- Objective: The final objective is an approximated upper-bound that estimates average dot products with mini-batch data, producing a simple training loss.The upper-bound requires the background-set probability, and the approximation uses mini-batch samples.
- Objective: The two final-loss terms enforce local prediction consistency and disperse predictions of potentially dissimilar mini-batch features.A decaying hyperparameter weakens the dispersing term during training because features gradually cluster and the approximation may be inaccurate.
- Relation to Existing Works: AaD relates domain adaptation, SFDA, and contrastive methods through discriminability and diversity, while differing from neighborhood clustering by considering similar and dissimilar features together.The comparison includes mutual-information, nuclear-norm, and neighborhood-clustering perspectives; marginal-entropy methods may presume uniform class distributions.
4 Experiments
Experiments evaluate AaD on closed-set, open-set, and partial-set source-free domain adaptation, including benchmark comparisons, ablations, runtime, and clustering analyses. AaD achieves strong benchmark performance while supporting unsupervised hyperparameter selection and efficient memory-bank variants.
- Quantitative Results: AaD achieves state-of-the-art source-free performance on Office-31 and VisDA, including a 2.1% improvement over NRC on VisDA.It achieves results similar to the more complex A2Net on Office-Home and comparable or better performance than source-present methods on Office-Home and VisDA.
- Ablation Analysis: Decision-boundary visualizations show that both objective terms are necessary and that decaying the second term improves training.The toy experiment uses a 30° rotation to instantiate the target-domain shift.
- Ablation Analysis: AaD is robust to the number of nearest neighbors, with a reasonable choice such as 3 working well across datasets.Too few neighbors can be noisy when misclassified, whereas too many may include samples from other categories; VisDA uses 5 for fair comparison with prior methods.
- Ablation Analysis: SND successfully selects the decay factor β without supervision, consistent with weakening the diversity term during training.The second term prevents collapse toward a limited set of categories, while its influence should decrease as features cluster.
- Runtime Analysis: A limited memory bank can reduce computation on very large datasets while retaining an efficient training strategy.The runtime analysis compares AaD and SHOT with different memory-bank buffer sizes.
- Generalization: The method extends to source-free open-set and partial-set domain adaptation, with much better balance between known-class and unknown accuracy in open-set evaluation.Open-set performance is reported using OS*, UNK, and their harmonic mean HOS.
5 Conclusion
The paper proposes AaD for source-free domain adaptation by clustering similar features and dispersing dissimilar predictions, using a simple upper-bound objective with discriminability and diversity terms. It reports state-of-the-art performance on several benchmarks and applicability to source-free open-set and partial-set domain adaptation.
- AaD encourages similar features to receive similar predictions and dissimilar features to receive dispersed predictions, jointly achieving feature clustering and cluster assignment.
- The proposed objective is replaced by an upper bound containing two simple terms.
- The paper relates domain adaptation, source-free domain adaptation, and contrastive learning methods through discriminability and diversity.
- AaD achieves state-of-the-art performance on several benchmarks and extends to source-free open-set and partial-set domain adaptation.
Checklist
The checklist reports that the paper addresses contribution and scope questions, provides reproducibility materials and training details, and does not report error bars, compute resources, or asset licenses. It also reports no discussion of potential negative societal impacts and no applicable human-subject or personally identifiable information disclosures.
- The checklist marks the paper’s contribution and scope description as adequate but says limitations were not discussed.
- The checklist reports no discussion of potential negative societal impacts.
- The authors report that code, data, and reproduction instructions are included, while error bars and compute details are not reported.
- Existing assets are cited, but their licenses are not mentioned.
- Human-subject materials and personally identifiable information disclosures are marked not applicable.