Research map / Label-efficient and robust learning / Noisy labels
Self-Filtering: A Noise-Aware Sample Selection for Label Noise with Confidence Penalization
Qi Wei, Haoliang Sun, Xiankai Lu, Yilong Yin
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 |
|---|---|---|---|
| replace small-loss selection with the fluctuation criterion (ablation) | Co-teaching · CIFAR-10 (symmetric 80%) | test accuracy (%): 34.6 -> 39.7 | Improved |
| replace small-loss selection with the fluctuation criterion (ablation) | JoCoR · CIFAR-10 (symmetric 80%) | test accuracy (%): 38.0 -> 41.6 | Improved |
| replace small-loss selection with the fluctuation criterion | DivideMix · CIFAR-10 and CIFAR-100 | test accuracy (%): 93.4 -> 94.1 (CIFAR-10 pair 40%); 93.0 -> 92.8 (CIFAR-10 symmetric 80%) | Mixed |
| replace small-loss selection with the fluctuation criterion | DivideMix · CIFAR-100 (pair 40%) | test accuracy (%): 71.0 -> 72.4 | Improved |
| add FixMatch semi-supervised learning on discarded samples (SFT+) | SFT · CIFAR-10 (instance 40%) | test accuracy (%): 89.97 -> 94.4 | Improved |
| add FixMatch semi-supervised learning on discarded samples (SFT+) | SFT · CIFAR-100 (instance 40%) | test accuracy (%): 69.91 -> 74.1 | Improved |
| use MixMatch instead of FixMatch | SFT+ · CIFAR-10 / CIFAR-100 noise settings | test accuracy (%): 92.9 -> 93.7 (pair 40%); 94.4 -> 94.1 (instance 40%) | Mixed |
| add warm-up and main-learning confidence regularizers R and L_CR (ablation) | SFT · 40% noise ratio | test accuracy (%) and F-score: 3.72% accuracy and 0.023 F-score decrease when removed | Improved |
| use fluctuation criterion instead of majority voting (ablation) | SFT · noisy CIFAR | test accuracy and F-score | Improved |
| use larger confidence threshold Gamma (0.2 to 0.8) (ablation) | SFT · CIFAR-10/100 | test accuracy (%) | No clear effect |
| use larger memory bank size T (ablation) | SFT · CIFAR-10/100 | test accuracy and F-score | Hurt |
| add MixMatch SSL (SFT+*) vs ELR+ on Clothing1M | SFT+* · Clothing1M | test accuracy (%): 0.27% improvement over ELR+ | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Me-Momentum — Me-Momentum: Extracting Hard Confident Examples From Noisily Labeled Data (ICCV 2021) · better on 7 of 7 shared results
- DivideMix — DivideMix: Learning with Noisy Labels as Semi-supervised Learning (ICLR 2020) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Combating Label Noise with a General Surrogate Model for Sample Selection (IJCV 2024 · similarity 0.89)
- Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples (NeurIPS 2025 · similarity 0.89)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.88)
- Label Noise: Ignorance Is Bliss (NeurIPS 2024 · similarity 0.88)
- Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources (ICML 2024 · similarity 0.87)
- Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.87)
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.
