Research map / Label-efficient and robust learning / Domain adaptation and generalization
Gradient-Based Sample Selection for Black-Box Universal 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 GSS, a black-box universal domain adaptation method that selects likely unknown-class samples via a gradient-derived Bayes criterion (no manual thresholds), combined with an open-set classifier, class-balancing losses, and self-training on target data using only the source model's interface.
Problem
Black-box universal domain adaptation (B2-UniDA): labelling unlabelled target data under unknown domain shift and category shift when only the query interface of a pre-trained source model is available, without source data or model weights.
Approach
A sample selection criterion T(x) derived from gradient descent on symmetric cross-entropy between an open-set classifier and a closed-set classifier (linked via Bayes' theorem) that separates known from unknown samples without thresholds, plus a balancing mechanism that adapts loss weights using estimated training status and DA type, and self-supervised prototype clustering on target data.
Key results
- In OPDA, GSS reaches 87.1% average H-score on Office vs 85.0% for UB2DA and 87.0% for UMAD.
- On OfficeHome OPDA, GSS averages 77.9% H-score, outperforming all listed UniDA, SF-UniDA and B2-UniDA methods.
- On VisDA OPDA, GSS achieves 43.5 H-score vs 25.5 for UB2DA; the paper reports GSS is about 2.8x UB2DA on VisDA in ODA and CDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GSS instead of UB2DA | B2-UniDA · Office OPDA (mean) | H-score: 85.0 -> 87.1 | Helped |
| use GSS instead of UB2DA | B2-UniDA · VisDA OPDA | H-score: 25.5 -> 43.5 | Helped |
| use GSS instead of UB2DA | B2-UniDA · OfficeHome OPDA (mean) | H-score: 69.9 -> 77.9 | Helped |
| use GSS instead of UB2DA | B2-UniDA · DomainNet OPDA (mean) | H-score: 46.8 -> 50.3 | Helped |
| add self-supervised clustering loss Lreg (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 0.0 -> 87.1; 0.0 -> 77.9; 0.0 -> 43.5; 48.8 -> 50.3 | Mixed |
| add known-sample self-training loss Lt-kn (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 84.9 -> 87.1; 76.4 -> 77.9; 23.1 -> 43.5; 40.1 -> 50.3 | Helped |
| add known-class probability correction loss L(1)p-kn (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 84.9 -> 87.1; 76.8 -> 77.9; 35.7 -> 43.5; 49.3 -> 50.3 | Helped |
| add unknown-class self-training loss L(1)p-un (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 0.0 -> 87.1; 1.8 -> 77.9; 0.0 -> 43.5; 13.3 -> 50.3 | Helped |
| add unknown-sample entropy self-training loss L(2)pse (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 83.2 -> 87.1; 73.7 -> 77.9; 33.5 -> 43.5; 49.2 -> 50.3 | Helped |
| add adaptive probability-estimation loss Lscrs (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 84.4 -> 87.1; 75.9 -> 77.9; 30.0 -> 43.5; 35.8 -> 50.3 | Helped |
| add multi-category balancing loss Lavr (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 84.7 -> 87.1; 72.9 -> 77.9; 27.8 -> 43.5; 12.7 -> 50.3 | Mixed |
| add full-batch balancing term L(1)full (ablation) | GSS · OPDA: Office, OfficeHome, VisDA, DomainNet | H-score: 87.1 -> 87.1; 77.9 -> 77.9; 43.5 -> 43.5; 33.6 -> 50.3 | Mixed |
Surprising findings
- Ablation shows removing the sample selection-related loss L(1)_p-un or omega_un drops DomainNet H-score to 13.3 and 0.0, indicating strong dependence on these components.
- Several reproduced baselines (UAN, CMU) required tuning and default hyperparameters gave poor results, suggesting baseline sensitivity.
Limitations
- Performance on VisDA and DomainNet is only comparable to best SF/UniDA methods in some settings rather than uniformly superior.
- Several hyperparameters (K_clu, e0, epoch thresholds) are set manually per dataset.
- Baseline UMAD results are taken raw from the original paper because its code is inaccessible.
Experimental setup
Datasets: Office OfficeHome VisDA DomainNet
Benchmarks: Office OfficeHome VisDA DomainNet
Metrics: H-score accuracy
Baselines: UB2DA USFDA UMAD UAN CMU DANCE DCC OVANet Source-only
Builds on: UB2DA DEPICT DEC OVANet UMAD
Closest papers by meaning
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.91)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.90)
- Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.90)
- ADU: Adaptive Detection of Unknown Categories in Black-Box Domain Adaptation (CVPR 2025 · similarity 0.90)
- LEAD: Learning Decomposition for Source-free Universal Domain Adaptation (CVPR 2024 · similarity 0.89)
- A Separation and Alignment Framework for Black-Box Domain Adaptation (AAAI 2024 · similarity 0.89)
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.