Research map / Label-efficient and robust learning / Noisy labels
Does label smoothing mitigate label 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 |
|---|---|---|---|
| Apply label smoothing rather than uncorrected cross-entropy | ResNet classifiers trained on symmetric noisy labels · CIFAR-10 and CIFAR-100 | test accuracy: 57.06±0.38 -> 60.70±0.28; 80.44±0.63 -> 83.95±0.18 | Helped |
| Apply label smoothing instead of the baseline | ResNet-v2-50 · ImageNet with 20% symmetric label noise | test accuracy: 70.86 -> 71.55 | Helped |
| Increase label-smoothing α from 0.0 to 0.2 (ablation) | ResNet-v2-50 · ImageNet with 20% symmetric label noise | test accuracy: 70.86 -> 71.55 | Helped |
| Increase α beyond the noise rate (ablation) | Label smoothing and loss-correction methods · CIFAR-10, CIFAR-100, and ImageNet with 20% symmetric noise | test accuracy: 70.86 -> 71.55 | Helped |
| Use forward correction rather than label smoothing | ResNet classifiers · CIFAR-100 with 20% symmetric label noise | test accuracy: 60.70±0.28 -> 61.29±0.38 | Helped |
| Use backward correction smoothing rather than the baseline | ResNet-32 · CIFAR-100 with 20% symmetric label noise | test accuracy: 57.06±0.38 -> 53.91±0.40 | Hurt |
| Increase label smoothing α from 0.0 to 0.2 (ablation) | ResNet-32 · CIFAR-100 train examples with flipped labels | accuracy on true labels: 39.92 -> 53.01 | Helped |
| Increase label smoothing α from 0.0 to 0.2 (ablation) | ResNet-32 · CIFAR-100 train examples with flipped labels | accuracy on observed noisy labels: 17.88 -> 8.32 | Hurt |
| Apply label smoothing to the teacher before distillation | ResNet-32 teacher and student · CIFAR-100 under label noise | student accuracy: 63.98±0.26 -> 64.48±0.25 | Helped |
| Apply label smoothing to the teacher before distillation | ResNet-32 teacher and student · CIFAR-10 under label noise | student accuracy: 80.44±0.64 -> 86.95±1.82 | Helped |
| Apply label smoothing to the student rather than the teacher (ablation) | ResNet-32 teacher and student · CIFAR-100 under label noise | student accuracy: 63.98±0.26 -> 63.83±0.28 | No clear effect |
| Apply label smoothing to the student rather than the teacher (ablation) | ResNet-32 teacher and student · CIFAR-10 under label noise | student accuracy: 80.44±0.64 -> 85.72±2.61 | Helped |
Closest papers by meaning
- Does Label Smoothing Help Deep Partial Label Learning? (ICML 2024 · similarity 0.91)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.89)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.89)
- NoiseGPT: Label Noise Detection and Rectification through Probability Curvature (NeurIPS 2024 · similarity 0.89)
- For Better or For Worse? Learning Minimum Variance Features With Label Augmentation (ICLR 2025 · similarity 0.88)
- Training Deep Neural Networks with Virtual Smoothing Classes (AAAI 2025 · 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.