Research map / Label-efficient and robust learning / Semi-supervised learning
Unknown-Aware Graph Regularization for Robust Semi-supervised Learning from Uncurated Data
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 unknown-aware graph regularization (UAG) that uses a multi-head uncertainty-based open-set classifier plus a contrastive graph loss over pseudo-label and pseudo-outlier graphs to improve open-set semi-supervised learning under correlated and uncorrelated outliers.
Problem
Semi-supervised learning degrades when unlabeled data contains out-of-class outliers, and existing open-set SSL methods either rely on feature similarity (failing on correlated outliers) or uncertainty in logits (which underperform).
Approach
Uses separate closed-set and open-set classifier heads so inlier and outlier objectives do not conflict, and builds a contrastive target graph combining known-class pseudo-label similarity with outlier-indicator agreement so that outliers form multiple clusters rather than one collapsed unknown class.
Key results
- On CIFAR-100 (100 labels/class, correlated outliers), UAG reduces error rate to 6.8% versus 9.5% for CoMatch-style baselines and raises AUROC to 96.4 versus 59.4 for FixMatch.
- On ImageNet-30 with 10% labels, UAG achieves 97.4 AUROC and the paper reports state-of-the-art error rates across protocols.
- On Semi-iNature-2021, UAG reaches 24.16% top-1 accuracy versus 19.17% for CoMatch.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add unknown-aware graph regularization to FixMatch (ablation) | CIFAR-100, 100 labels/class, uncorrelated · CIFAR-100 uncorrelated | error rate: 29.4 -> 23.6 | Helped |
| Add unknown-aware graph regularization to FixMatch (ablation) | CIFAR-100, 100 labels/class, uncorrelated · CIFAR-100 uncorrelated | AUROC: 66.7 -> 91.8 | Helped |
| Add unknown-aware graph regularization to FixMatch (ablation) | CIFAR-100, 100 labels/class, correlated · CIFAR-100 correlated | error rate: 28.8 -> 23.8 | Helped |
| Add unknown-aware graph regularization to FixMatch (ablation) | CIFAR-100, 100 labels/class, correlated · CIFAR-100 correlated | AUROC: 62.6 -> 81.4 | Helped |
| Add entropy maximization (EMX) (ablation) | FixMatch · CIFAR-100 uncorrelated | AUROC: 66.7 -> 82.2 | Helped |
| Add entropy maximization (EMX) (ablation) | FixMatch · CIFAR-100 uncorrelated | error rate: 29.4 -> 29.3 | No clear effect |
| Add multi-head training (MHT) (ablation) | FixMatch with EMX · CIFAR-100 uncorrelated | error rate: 29.3 -> 27.5 | Helped |
| Add multi-head training (MHT) (ablation) | FixMatch with EMX · CIFAR-100 uncorrelated | AUROC: 82.2 -> 89.1 | Helped |
| Add the pseudo-label graph (PLG) (ablation) | FixMatch with multi-head training · CIFAR-100 uncorrelated | error rate: 27.5 -> 25.8 | Helped |
| Add the pseudo-label graph (PLG) (ablation) | FixMatch with multi-head training · CIFAR-100 uncorrelated | AUROC: 89.1 -> 63.9 | Hurt |
| Add the pseudo-outlier graph (POG) (ablation) | FixMatch with multi-head training and PLG · CIFAR-100 uncorrelated | AUROC: 63.9 -> 91.8 | Helped |
| Add the pseudo-outlier graph (POG) (ablation) | FixMatch with multi-head training and PLG · CIFAR-100 uncorrelated | error rate: 25.8 -> 23.6 | No clear effect |
Surprising findings
- Uncertainty-based OSSL methods with a single classifier performed worse than similarity-based methods in the paper's experiments, and conventional SSL shows low inlier-outlier discrimination yet high discrimination among outlier classes in embedding space.
Limitations
- Experiments on ImageNet use a 30-class subset (ImageNet-30) due to computational constraints rather than full ImageNet.
- Hyperparameters such as graph threshold and temperature are fixed across experiments for simplicity rather than tuned per dataset.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet-30 Semi-iNature-2021
Benchmarks: CIFAR-10 CIFAR-100 ImageNet-30 Semi-iNature-2021
Metrics: error rate AUROC top-1 accuracy top-5 accuracy
Baselines: Labeled Only FixMatch CoMatch MTC OpenMatch OSP Ent-Max UDM
Builds on: FixMatch CoMatch OpenMatch Entropy maximization for OSSL (Sun and Wang 2022) Uniform distribution matching (He et al. 2022)
Closest papers by meaning
- Out-Of-Domain Unlabeled Data Improves Generalization (ICLR 2024 · similarity 0.87)
- ROG_PL: Robust Open-Set Graph Learning via Region-Based Prototype Learning (AAAI 2024 · similarity 0.86)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.86)
- Unlocking the Potential of Unlabeled Data in Semi-Supervised Domain Generalization (CVPR 2025 · similarity 0.86)
- Semi-supervised Graph Anomaly Detection via Robust Homophily Learning (NeurIPS 2025 · similarity 0.86)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · 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.