Research map / Label-efficient and robust learning / Semi-supervised learning
CGMatch: A Different Perspective of Semi-supervised Learning
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 CGMatch, a semi-supervised learning method that uses a new Count-Gap metric (gap between top two pseudo-label prediction counts) with confidence to split unlabeled data into easy, ambiguous, and hard subsets, applying CE and GCE losses respectively.
Problem
Semi-supervised image classification with very scarce labels, where confidence-based pseudo-label selection (as in FixMatch-style X-Match methods) struggles to identify useful unlabeled examples and filter noisy ones early in training.
Approach
Introduces Count-Gap (CG), the difference between the counts of the most and second-most frequently predicted pseudo-labels for each unlabeled sample over training, as a label-free analogue of data-map variability. Combined with confidence in a fine-grained dynamic selection (FDS) scheme, unlabeled data is split into easy-to-learn (CE loss), ambiguous (GCE loss), and hard-to-learn (discarded) subs
Key results
- CGMatch achieves best results in 8 of 12 settings on CIFAR10, CIFAR100, SVHN, and STL10 with 4/10/25 labels per class, e.g., 4.87% error on CIFAR10-4 and 47.55% on CIFAR100-4.
- On CIFAR100-4, CGMatch improves over FlexMatch by 2.60 percentage points.
- CGMatch underperforms SoftMatch on STL10-10 and STL10-25 (16.23% and 9.48% vs. SoftMatch's 11.59% and 8.01%... see table), attributed to the open-set nature of STL10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CGMatch instead of FlexMatch | semi-supervised image classification · CIFAR100-4 | top-1 error rate (%): 50.15±1.51 -> 47.55±2.48 | Helped |
| Use CGMatch instead of SoftMatch | semi-supervised image classification · STL10-10 and STL10-25 | top-1 error rate (%): 11.59±2.60 -> 16.23±2.39; 8.01±0.32 -> 9.48±0.60 | Hurt |
| Replace Self-Adaptive Thresholding with Global Thresholding (ablation) | CGMatch · CIFAR10, CIFAR100, SVHN, and STL10 | top-1 error rate (%): CIFAR10-4: 5.19±0.22 -> 4.87±0.18; SVHN-25: 5.57±1.72 -> 2.13±0.05; STL10-10: 16.38±2.19 -> 16.23±2.39; CIFAR10-25: 4.83 | Mixed |
| Include ambiguous samples with FDS and train them using GCE loss | CGMatch · STL10-4 | iterations to reach evaluation accuracy 0.7: 135k iterations | Helped |
Surprising findings
- CGMatch lags SoftMatch on STL10-10 and STL10-25, which the authors attribute to STL10 being open-set.
Limitations
- Underperforms on STL10 with more labels, likely due to open-set unlabeled data; open-set SSL not addressed.
- Hard-to-learn samples are discarded rather than exploited.
Experimental setup
Datasets: CIFAR10 CIFAR100 SVHN STL10
Benchmarks: CIFAR10 CIFAR100 SVHN STL10
Metrics: top-1 error rate expected calibration error utilization ratio
Baselines: PseudoLabel MeanTeacher MixMatch ReMixMatch FixMatch FlexMatch SoftMatch FreeMatch FullFlex
Builds on: FixMatch FreeMatch SoftMatch FlexMatch data maps
Closest papers by meaning
- Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.90)
- InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · similarity 0.90)
- Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection (NeurIPS 2024 · similarity 0.90)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.90)
- Towards Realistic Model Selection for Semi-supervised Learning (ICML 2024 · similarity 0.89)
- Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · similarity 0.89)
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.