Research map / Label-efficient and robust learning / Domain adaptation and generalization
A Separation and Alignment Framework for Black-Box Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SEAL, a black-box domain adaptation framework that separates target samples into well- and under-adapted subsets and aligns under-adapted samples via a graph contrastive objective plus nearest-centroid classification, using only soft predictions from black-box source models.
Problem
Black-box domain adaptation: learning a classifier for an unlabeled target domain when only black-box source predictors (soft outputs) are available, without source data or white-box weights; prior methods rely on a fixed budget of easy samples and ignore target-domain structure.
Approach
Dynamically separate target samples into well-adapted (loss-ranked easy samples plus high-confidence samples chosen by class-wise adaptive thresholds) and under-adapted sets, then align under-adapted samples to local neighbors and well-adapted peers via a graph contrastive learning loss between a locality label graph and a feature graph, augmented by a nearest-centroid objective.
Key results
- On VisDA, SEAL improves averaged class-wise accuracy to 89.2%, a 4.1% gain over the best baseline.
- On Office-Home, SEAL reaches 74.5% average accuracy, improving BETA by 2.4%.
- On DomainNet, SEAL averages 55.8% versus 54.5% for BETA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SEAL instead of BETA | black-box domain adaptation · Office-Home | accuracy: 72.1 -> 74.5 | Helped |
| apply SEAL instead of BETA | black-box domain adaptation · Office | accuracy: 88.2 -> 88.8 | Helped |
| apply SEAL instead of BETA | black-box domain adaptation · VisDA | class-average accuracy: 85.1 -> 89.2 | Helped |
| apply SEAL instead of BETA | black-box domain adaptation · DomainNet | accuracy: 54.5 -> 55.8 | Helped |
| apply SEAL instead of DINE | multi-source black-box domain adaptation · Office-Home | accuracy: 77.1 -> 78.8 | Helped |
| add domain separation (ablation) | SEAL · VisDA | accuracy: 84.3 -> 86.4 | Helped |
| add model-confident sample selection (ablation) | SEAL · VisDA | class-average accuracy: 87.5 -> 89.2 | Helped |
| add graph contrastive learning loss (ablation) | SEAL · VisDA | class-average accuracy: 85.3 -> 89.2 | Helped |
| add neighborhood alignment (ablation) | SEAL · VisDA | class-average accuracy: 87.9 -> 89.2 | Helped |
| add nearest-centroid classification (ablation) | SEAL · VisDA | class-average accuracy: 87.1 -> 89.2 | Helped |
| use a positive scaling parameter γ instead of γ = 0.0 (ablation) | SEAL model-confident sample selection · VisDA | accuracy | Helped |
Limitations
- Relies on a shared subdomain assumption between source and target.
- Several hyperparameters (tau, gamma, eta, k, M) are tuned per dataset, with gamma set differently for each dataset.
Experimental setup
Datasets: Office Office-Home VisDA DomainNet
Benchmarks: Office Office-Home VisDA DomainNet
Metrics: top-1 accuracy average per-class accuracy
Baselines: Source NLL-OT NLL-KL HD-SHOT SD-SHOT DINE BETA
Builds on: DINE BETA NLL-KL SHOT supervised contrastive learning
Closest papers by meaning
- Reviewing the Forgotten Classes for Domain Adaptation of Black-Box Predictors (AAAI 2024 · similarity 0.90)
- ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation (CVPR 2025 · similarity 0.89)
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.89)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.87)
- Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.86)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.