Research map / Label-efficient and robust learning / Domain adaptation and generalization
LEAD: Learning Decomposition for Source-free 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 LEAD, which uses SVD-based orthogonal feature decomposition and instance-level adaptive decision boundaries to identify target-private data in source-free universal domain adaptation without thresholds or iterative clustering.
Problem
Source-free universal domain adaptation (SF-UniDA) requires separating covariate-shifted target samples into common and target-private unknown classes without source data, but existing methods rely on hand-crafted thresholds or slow iterative clustering.
Approach
Orthogonally decompose target features via SVD of the frozen classifier weights into source-known and source-unknown components; the source-unknown norm indicates private data, and a GMM plus a per-instance common score (distances to target prototypes and source anchors) yields instance-level decision boundaries for pseudo-labeling.
Key results
- In OPDA on VisDA, LEAD reaches 76.6% H-score, surpassing GLC by 3.5%.
- LEAD improves UMAD's OPDA H-score on Office-Home from 70.1% to 78.0%.
- Runtime to derive decision boundaries on DomainNet drops from 897.65 s (GLC) to 0.29 s (LEAD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature decomposition regularization Lreg (ablation) | LEAD with Lce · Office-31, Office-Home, and VisDA OPDA | H-score (%): 82.0 -> 84.5; 73.6 -> 74.5; 61.5 -> 67.3 | Helped |
| add feature consensus regularization Lcon (ablation) | LEAD with Lce · Office-31, Office-Home, and VisDA OPDA | H-score (%): 82.0 -> 87.2; 73.6 -> 74.5; 61.5 -> 75.7 | Helped |
| add Lcon to LEAD with Lce and Lreg (ablation) | LEAD · Office-31 and VisDA OPDA | H-score (%): 84.5 -> 87.8; 67.3 -> 76.6 | Helped |
| add Lcon to LEAD with Lce and Lreg (ablation) | LEAD · Office-Home OPDA | H-score (%): 74.5 -> 75.0 | Helped |
| use instance-level decision boundaries instead of a global threshold (ablation) | LEAD · Office-Home OPDA | H-score (%): 63.1% vs 75.0% | Helped |
| use entropy instead of the source-unknown feature projection as the private-data indicator (ablation) | LEAD · Office-31, Office-Home, and VisDA | H-score (%) | Hurt |
| integrate LEAD | UMAD · Office-Home OPDA | H-score (%): 70.1% to 78.0% (+7.9) | Helped |
| integrate LEAD | GLC · VisDA OPDA | H-score (%): 73.1% to 76.8% (+3.7) | Helped |
| integrate LEAD | UMAD · VisDA OPDA | H-score (%): 58.3% to 67.2% (+8.9) | Helped |
| integrate LEAD | UMAD · Office-31 PDA | accuracy (%): +4.1% | Helped |
| integrate LEAD | GLC · Office-Home OPDA | H-score (%): +0.6 | Helped |
| integrate LEAD | GLC · VisDA PDA | accuracy (%): -0.7% | Hurt |
Surprising findings
- Adding LEAD to GLC slightly degrades PDA accuracy on VisDA (-0.7%), which the authors attribute to possible gradient conflict.
- Entropy as the privacy indicator performs worse than the source-unknown feature norm within the same framework.
Limitations
- Integration with GLC is not uniformly beneficial (slight PDA degradation on VisDA).
- Relies on the bimodal distribution of source-unknown feature norms being modeled by a two-component GMM.
Experimental setup
Datasets: Office-31 Office-Home VisDA DomainNet
Benchmarks: Office-31 Office-Home VisDA DomainNet
Metrics: H-score accuracy runtime
Baselines: CMU DANCE DCC OVANet GATE UniOT SHOT-O UMAD GLC
Builds on: GLC UMAD SHOT Universal domain adaptation
Closest papers by meaning
- Gradient-Based Sample Selection for Black-Box Universal Domain Adaptation (AAAI 2025 · similarity 0.89)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.89)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.88)
- Distinguish Then Exploit: Source-free Open Set Domain Adaptation via Weight Barcode Estimation and Sparse Label Assignment (CVPR 2025 · similarity 0.88)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.87)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.87)
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.