Research map / Label-efficient and robust learning / Semi-supervised learning
Unlocking the Potential of Unlabeled Data in Semi-Supervised Domain Generalization
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 UPCSC, a plug-and-play method that uses unconfident-unlabeled samples in semi-supervised domain generalization via unlabeled proxy-based contrastive learning and surrogate class learning.
Problem
Semi-supervised domain generalization (SSDG) methods discard unlabeled samples whose predictions fall below a confidence threshold, underusing the available unlabeled data under domain shift.
Approach
Treats unconfident-unlabeled samples as extra negative pairs against classes outside their candidate set (UPC) and builds a confidence-weighted surrogate class proxy from candidate classes as a positive pair (SC), requiring no domain labels and attaching to existing SSL-based SSDG baselines.
Key results
- With FixMatch as baseline, UPCSC raises PACS accuracy from 76.8 to 79.6 (10 labels per class) and DigitsDG from 75.1 to 80.7.
- With StyleMatch, UPCSC reaches 82.2% on DigitsDG (10 labels per class), a +3.8 gain over the baseline.
- Ablation on PACS and OfficeHome (10 labels per class) shows UPC and SC together give +2.8 and +0.9 points over FixMatch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add UPCSC to FixMatch | FixMatch · PACS, 10 labels per class | accuracy: 76.8 -> 79.6 | Helped |
| Add UPCSC to FixMatch | FixMatch · OfficeHome, 10 labels per class | accuracy: 57.7 -> 58.6 | Helped |
| Add UPCSC to FixMatch | FixMatch · DigitsDG, 10 labels per class | accuracy: 75.1 -> 80.7 | Helped |
| Add UPCSC to FixMatch | FixMatch · miniDomainNet, 10 labels per class | accuracy: 54.5 -> 56.0 | Helped |
| Add UPCSC to FixMatch | FixMatch · PACS, 5 labels per class | accuracy: 73.6 -> 78.9 | Helped |
| Add UPCSC to FixMatch | FixMatch · OfficeHome, 5 labels per class | accuracy: 55.0 -> 56.1 | Helped |
| Add UPCSC to FixMatch | FixMatch · DigitsDG, 5 labels per class | accuracy: 64.7 -> 75.2 | Helped |
| Add UPCSC to FixMatch | FixMatch · miniDomainNet, 5 labels per class | accuracy: 51.6 -> 52.7 | Helped |
| Add UPC to FixMatch without SC (ablation) | FixMatch · PACS, 10 labels per class | accuracy: 76.8 -> 79.2 | Helped |
| Add UPC to FixMatch without SC (ablation) | FixMatch · OfficeHome, 10 labels per class | accuracy: 57.7 -> 58.4 | Helped |
| Add SC to FixMatch without UPC (ablation) | FixMatch · PACS, 10 labels per class | accuracy: 76.8 -> 77.0 | Helped |
| Add SC to FixMatch without UPC (ablation) | FixMatch · OfficeHome, 10 labels per class | accuracy: 57.7 -> 58.4 | Helped |
Surprising findings
- SC alone gives only +0.2 on PACS, while UPC alone gives +2.4, indicating the contrastive-negative module carries most of the PACS gain.
Limitations
- Gains on some benchmarks are small (e.g., OfficeHome +0.2 to +1.4 points) and some variants are within standard deviation.
- Evaluated only with ResNet18 backbone and 20 training epochs.
Experimental setup
Datasets: PACS OfficeHome DigitsDG miniDomainNet
Benchmarks: PACS OfficeHome DigitsDG miniDomainNet
Metrics: top-1 accuracy
Baselines: ERM MeanTeacher FreeMatch FixMatch StyleMatch FBCSA DGWM
Builds on: FixMatch proxy-based contrastive learning (PCL) StyleMatch FreeMatch
Closest papers by meaning
- Towards Generalizing to Unseen Domains with Few Labels (CVPR 2024 · similarity 0.92)
- Out-Of-Domain Unlabeled Data Improves Generalization (ICLR 2024 · similarity 0.91)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.89)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.89)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 2024 · similarity 0.89)
- Understanding Domain Generalization: A Noise Robustness Perspective (ICLR 2024 · similarity 0.88)
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.