Research map / Label-efficient and robust learning / Semi-supervised learning
FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreeMatch instead of FlexMatch | SSL · CIFAR-10 (10 labels) | error rate: 13.85 ± 12.04 -> 8.07 ± 4.24 | Helped |
| Use FreeMatch instead of FlexMatch | SSL · STL-10 (40 labels) | error rate: 29.15 ± 4.16 -> 15.56 ± 0.55 | Helped |
| Use FreeMatch instead of FlexMatch | SSL · ImageNet (100 labels per class) | Top-1 error rate: 41.85 -> 40.57 | Helped |
| Use FreeMatch instead of FlexMatch | SSL · CIFAR-100 (2500 labels) | error rate: 26.49 ± 0.20 -> 26.47 ± 0.20 | Hurt |
| Combine global and local self-adaptive thresholds (ablation) | FreeMatch thresholding · CIFAR-10 (40 labels) | error rate: 6.06 -> 4.92 | Helped |
| Add local self-adaptive threshold modulation (ablation) | FreeMatch thresholding · CIFAR-10 (40 labels) | error rate: 6.06 -> 4.92 | Helped |
| Replace local self-adaptive modulation with FlexMatch CPL (ablation) | global self-adaptive threshold · CIFAR-10 (40 labels) | error rate: 6.06 -> 8.40 | Hurt |
| Add SAF (ablation) | FreeMatch · CIFAR-10 (10 labels) | error rate: 10.37 -> 8.07 | Helped |
| Use original fairness objective instead of SAF (ablation) | FreeMatch · CIFAR-10 (10 labels) | error rate: 10.37 -> 9.57 | Helped |
| Add histogram normalization to the fairness objective (ablation) | original fairness objective · CIFAR-10 (10 labels) | error rate: 9.57 -> 12.07 | Hurt |
| Vary thresholding EMA decay among tested values (ablation) | FreeMatch · CIFAR-10 (40 labels) | error rate: 4.94, 4.92, 4.90, 5.03 | No clear effect |
| Add SAF (ablation) | FlexMatch · CIFAR-10 (10 labels) | error rate: 13.85 ± 12.04 -> 12.60 ± 8.16 | Helped |
Closest papers by meaning
- Class-Imbalanced Semi-Supervised Learning with Adaptive Thresholding (ICML 2022 · similarity 0.90)
- OwMatch: Conditional Self-Labeling with Consistency for Open-World Semi-Supervised Learning (NeurIPS 2024 · similarity 0.89)
- RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · similarity 0.89)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.87)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.86)
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.