Research map / Label-efficient and robust learning / Semi-supervised learning
Towards Generalizing to Unseen Domains with Few Labels
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 feature-based conformity and a semantics alignment loss as plug-and-play additions to SSL-based baselines (FixMatch, FlexMatch, FreeMatch, StyleMatch) for semi-supervised domain generalization on five DG benchmarks.
Problem
Semi-supervised domain generalization (SSDG): learning a model that generalizes to unseen domains when only a few labels are available across multiple source domains, with the rest unlabeled. Existing DG methods ignore unlabeled data, and SSL methods trail fully supervised DG.
Approach
Align same-domain and random different-domain posteriors computed from cosine similarities between unlabeled features and domain-aware labeled class prototypes with the pseudo-label (feature-based conformity), and add a semantics alignment loss that pulls features toward their assigned prototype and pushes them from hard non-assigned prototypes within the domain.
Key results
- With FixMatch on PACS (5 labels per class), the method improves average accuracy from 73.4 to 77.3 (ablation table).
- StyleMatch+Ours reaches 80.5% on PACS with 10 labels per class, versus 79.4% for StyleMatch alone.
- FixMatch+Ours gains 4.0% on DigitsDG (10 labels) and 6.0% (5 labels), and over 4% on TerraIncognita in both settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FBC and SA to StyleMatch | StyleMatch · PACS, 10 labels per class | accuracy (%): 79.4 -> 80.5 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · OfficeHome, 10 labels per class | accuracy (%): 59.7 -> 60.3 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · VLCS, 10 labels per class | accuracy (%): 73.5 -> 74.2 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · DigitsDG, 10 labels per class | accuracy (%): 65.9 -> 67.7 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · TerraIncognita, 10 labels per class | accuracy (%): 29.9 -> 32.5 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · PACS, 5 labels per class | accuracy (%): 78.4 -> 79.3 | Helped |
| Add FBC and SA to StyleMatch | StyleMatch · DigitsDG, 5 labels per class | accuracy (%): 55.7 -> 58.7 | Helped |
| Add same-domain FBC (ablation) | FixMatch · PACS, 5 labels per class | accuracy (%): 73.4 -> 76.0 | Helped |
| Add different-domain FBC (ablation) | FixMatch · PACS, 5 labels per class | accuracy (%): 73.4 -> 74.9 | Helped |
| Add FBC (ablation) | FixMatch · PACS, 5 labels per class | accuracy (%): 73.4 -> 76.7 | Helped |
| Add SA (ablation) | FixMatch · PACS, 5 labels per class | accuracy (%): 73.4 -> 74.8 | Helped |
| Add FBC and same-domain SA (ablation) | FixMatch · PACS, 5 labels per class | accuracy (%): 73.4 -> 77.0 | Helped |
Surprising findings
- StyleMatch fails to improve over FixMatch on datasets without style shifts (DigitsDG, TerraIncognita), while the proposed method gives large gains there.
- Several SSL baselines (FlexMatch, FreeMatch) underperform ERM on some datasets such as PACS in the 5-label setting and on OfficeHome.
Limitations
- May require significant adjustments to handle semi-supervised single-source domain generalization.
- Gains on some datasets are modest (e.g., OfficeHome and VLCS with FlexMatch show drops).
Experimental setup
Datasets: PACS OfficeHome VLCS DigitsDG TerraIncognita
Benchmarks: PACS OfficeHome VLCS DigitsDG TerraIncognita
Metrics: top-1 accuracy averaged over 5 trials
Baselines: ERM EntMin MeanTeacher FlexMatch FreeMatch FixMatch StyleMatch
Builds on: FixMatch StyleMatch FlexMatch FreeMatch
Closest papers by meaning
- Unlocking the Potential of Unlabeled Data in Semi-Supervised Domain Generalization (CVPR 2025 · similarity 0.92)
- Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization (NeurIPS 2025 · similarity 0.91)
- Reasoning-Driven Multimodal LLM for Domain Generalization (ICLR 2026 · similarity 0.89)
- Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.89)
- Rethinking Multi-domain Generalization with A General Learning Objective (CVPR 2024 · similarity 0.88)
- Robust Domain Generalization under Divergent Marginal and Conditional Distributions (NeurIPS 2026 · 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.