Research map / Label-efficient and robust learning / Noisy labels
Probabilistic End-To-End Noise Correction for Learning With Noisy Labels
Check 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 PENCIL instead of Trunc L_q | CIFAR-100, symmetric noise · CIFAR-100, symmetric 20%, 40% and 60% noise | accuracy: 67.61±0.18 -> 73.86±0.34; 62.64±0.33 -> 69.12±0.62; 54.04±0.56 -> 57.79±3.86 | Helped |
| Use PENCIL instead of Trunc L_q | CIFAR-100, symmetric 80% noise · CIFAR-100, symmetric 80% noise | accuracy: 29.60±0.51 -> fail | Hurt |
| Use PENCIL instead of the strongest baseline | CIFAR-100, asymmetric noise · CIFAR-100, asymmetric 10%, 20% and 30% noise | accuracy: 71.05±0.30 -> 75.93±0.20; 71.08±0.22 -> 74.70±0.56; 70.76±0.26 -> 72.52±0.38 | Helped |
| Use PENCIL instead of Forward T | CIFAR-100 · CIFAR-100, asymmetric 40% noise | accuracy: 70.82±0.45 -> 63.61±0.23 | Hurt |
| Use PENCIL instead of Tanaka et al. | CIFAR-10, symmetric noise · CIFAR-10, symmetric 10%, 30%, 50%, 70% and 90% noise | accuracy: 93.23 -> 93.28; 91.22 -> 92.24; 88.51 -> 90.36; 84.59 -> 87.18; 53.49 -> 60.80 | Helped |
| Use PENCIL instead of Tanaka et al. | CIFAR-10, asymmetric noise · CIFAR-10, asymmetric 10%, 20%, 30% and 50% noise | accuracy: 92.64 -> 93.04; 91.92 -> 92.43; 91.18 -> 91.80; 68.35 -> 80.06 | Helped |
| Use PENCIL instead of Tanaka et al. | CIFAR-10, asymmetric 40% noise · CIFAR-10, asymmetric 40% noise | accuracy: 91.55 -> 91.16 | Hurt |
| Use PENCIL instead of cross-entropy training | Clothing1M · Clothing1M | accuracy: 72.16 -> 73.49 | Helped |
| Use PENCIL with λ=3000 and α=0 instead of cross-entropy | CUB-200 · CUB-200 | accuracy: 81.93 -> 82.64 | Helped |
| Vary λ and α across the reported PENCIL settings (ablation) | CUB-200 · CUB-200 | accuracy: 81.93; 81.91; 81.84; 82.64; 82.09; 82.21; 82.22 | Mixed |
| Use probabilistic label estimates updated end-to-end instead of fixed noisy labels | PENCIL · CIFAR-10, 70% symmetric noise | correct training labels: about 16000 -> about 39000 | Helped |
Closest papers by meaning
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.87)
- Augmentation Strategies for Learning With Noisy Labels (CVPR 2021 · similarity 0.87)
- Making Deep Neural Networks Robust to Label Noise: A Loss Correction Approach (CVPR 2017 · similarity 0.87)
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.87)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.86)
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise (NeurIPS 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.