Research map / Label-efficient and robust learning / Domain adaptation and generalization
Cycle Self-Refinement for Multi-Source 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 CSR, a cycle self-refinement method for multi-source domain adaptation that aggregates source-specific networks' predictions per target instance to pseudo-label and refine a domain-ensemble network, which in turn guides the source-specific networks via conditional feature alignment.
Problem
Multi-source domain adaptation transfers knowledge from several labeled source domains to an unlabeled target domain, but domain-level source weighting ignores dominant transferable knowledge in weaker source domains for individual target samples.
Approach
Instance-level ensemble weights each source-specific network per target sample using prediction confidence, cluster-based prediction diversity, and ensemble diversity; high-confidence ensemble pseudo-labels train a domain-ensemble network, and that network's pseudo-labels drive conditional feature alignment for the source-specific networks in a cycle.
Key results
- CSR reaches 92.7% average accuracy on Office-31, the best among compared methods (PTMDA 91.7%).
- CSR achieves 80.1% average accuracy on Office-Home versus 79.4% for WADN.
- CSR achieves 52.9% average accuracy on DomainNet versus 51.3% for DRT+ST.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Lpsd to Lcls (ablation) | CSR · Office-Home | accuracy: 72.4 -> 78.8 | Helped |
| Add Lalign to Lcls + Lpsd (ablation) | CSR · Office-Home | accuracy: 78.8 -> 80.1 | Helped |
| Add cic to the instance-level ensemble strategy (ablation) | CSR · DomainNet | accuracy: 56.5 -> 57.2 | Helped |
| Add ccd to the instance-level ensemble strategy (ablation) | CSR · DomainNet | accuracy: 56.9 -> 57.2 | Helped |
| Add rd to the instance-level ensemble strategy (ablation) | CSR · DomainNet | accuracy: 57.0 -> 57.2 | Helped |
| Replace average weighting with instance-level ensemble (ablation) | CSR · DomainNet | accuracy: 56.1 -> 57.2 | Helped |
Surprising findings
- On Office-31, DCA's →A accuracy (55.1%) is far below other multi-source methods, and CSR is not best on every task (e.g., →Inf on DomainNet is 28.1% vs 31.6% for DRT+ST).
Limitations
- Optimal threshold tau differs substantially between Office-Home and DomainNet, requiring dataset-dependent tuning.
- Does not win on every task; some tasks show lower accuracy than baselines (e.g., DomainNet →Inf).
Experimental setup
Datasets: Office-31 Office-Home DomainNet
Benchmarks: Office-31 Office-Home DomainNet
Metrics: classification accuracy
Baselines: DANN MCD MFSAN DCA SImpAI SSG PTMDA WADN BDT MRF-MSDA DRT+ST
Builds on: Margin Disparity Discrepancy FixMatch MFSAN DANN
Closest papers by meaning
- Dynamic Domain Adaptation for Efficient Inference (CVPR 2021 · similarity 0.88)
- Dynamic Transfer for Multi-Source Domain Adaptation (CVPR 2021 · similarity 0.88)
- Agile Multi-Source-Free Domain Adaptation (AAAI 2024 · similarity 0.87)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · similarity 0.85)
- Unveiling the Unknown: Unleashing the Power of Unknown to Known in Open-Set Source-Free Domain Adaptation (CVPR 2024 · similarity 0.85)
- MLNet: Mutual Learning Network with Neighborhood Invariance for Universal Domain Adaptation (AAAI 2024 · similarity 0.85)
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.