Research map / Label-efficient and robust learning / Semi-supervised learning
Confidence Estimation Using Unlabeled Data
Chen Li, Xiaoling Hu, Chao Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| consistency ranking loss (plus correctness ranking loss) added to training | PreAct-ResNet110 on CIFAR-10 · CIFAR-10 with 2500/5000/10000 labels | AURC: e.g. 109.78 -> 81.50 at 2500 labels; 37.46 -> 28.39 at 10000 | Improved |
| consistency ranking loss (plus correctness ranking loss) added to training | DenseNetBC on CIFAR-100 · CIFAR-100 with 2500/5000/10000 labels | AURC: 214.17 -> 168.57 at 10000 labels | Improved |
| consistency ranking loss added to segmentation training | UNet ResNet34 on ISIC2017 · ISIC2017 1/16 (125) labels | AURC: 22.14 (CRL) -> 14.11 | Improved |
| consistency ranking loss added to segmentation training | UNet ResNet34 on ISIC2017 · ISIC2017 full (2000) labels | AURC: CRL 10.17 vs Ours 11.18 | Hurt |
| removing correctness ranking loss (ablation) | full method on CIFAR-10 (5000 labels) · CIFAR-10 5000 labels | AURC: 47.12 without vs 44.43 with correctness loss | Improved |
| min-max normalization of consistency per labeled/unlabeled batch separately vs unified normalization (ablation) | consistency targets · CIFAR-10 5000 labels | AURC: 48.03 (unified) vs 44.43 (separate) | Improved |
| ranking loss instead of point-wise L1 loss on consistency (ablation) | confidence output · CIFAR-10 5000 labels | AURC: 76.73 (L1) vs 44.43 (ranking) | Improved |
| lambda_2 perturbation for consistency ranking loss weight (ablation) | segmentation on ISIC2017 · ISIC2017 | confidence estimation metrics | No clear effect |
| consistency ranking loss on CIFAR-10 minibatch pairs | maximum softmax output · CIFAR-10 unlabeled training set | E-AURC: 29.07 (Softmax) -> 19.35 (Softmax ours) | Improved |
Closest papers by meaning
- Improving the Reliability for Confidence Estimation (ECCV 2022 · similarity 0.88)
- Towards More Reliable Confidence Estimation (TPAMI 2023 · similarity 0.87)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.86)
- Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification (ICML 2025 · similarity 0.85)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.85)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.84)
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.
