Research map / Label-efficient and robust learning / Noisy labels
Using Trusted Data to Train Deep Networks on Labels Corrupted by Severe Noise
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 GLC's trusted-set corruption-matrix estimate and corrected training instead of the strongest reported baseline | Label-noise classification · MNIST, CIFAR-10, CIFAR-100, SST2, IMDB, Twitter POS | AUC error (%): MNIST: 11.5 -> 5.0; CIFAR-10: 14.2 -> 6.9; CIFAR-100: 37.5 -> 30.2; SST2: 24.4 -> 23.3; IMDB: 22.5 -> 22.0; Twitter: 24. | Helped |
| Use GLC instead of the strongest reported baseline | Label-noise classification · SVHN, uniform/flip noise | AUC error (%): 16.1 -> 16.4 | Hurt |
| Use GLC rather than Forward on Clothing1M | ResNeXt-50 fine-tuning · Clothing1M | accuracy: 79.03% -> 80.67% | Helped |
| Add training on trusted labels to Forward loss correction (ablation) | Forward loss correction · Several datasets and corruption settings | AUC error (%) | Mixed |
| Use the true corruption matrix instead of the estimated matrix (ablation) | GLC on CIFAR-100 with flip corruption and trusted fraction 5/100 · CIFAR-100 | AUC error (%): +0.96 percentage points | Helped |
| Calibrate the noisy-label classifier's softmax probabilities with temperature scaling (ablation) | GLC corruption-matrix estimation · CIFAR-100, flip corruption, trusted fraction 5/100 | AUC error (%) | No clear effect |
| Incorporate label base rates into the corruption-matrix estimate (ablation) | GLC corruption-matrix estimation · CIFAR-100, flip corruption, trusted fraction 5/100 | AUC error (%) | No clear effect |
Closest papers by meaning
- Distilling Effective Supervision From Severe Label Noise (CVPR 2020 · similarity 0.93)
- Error-Bounded Correction of Noisy Labels (ICML 2020 · similarity 0.92)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.89)
- Combating Semantic Contamination in Learning with Label Noise (AAAI 2025 · similarity 0.88)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.88)
- Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement (CVPR 2024 · similarity 0.88)
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.