Research map / Label-efficient and robust learning / Noisy labels
Deep Self-Learning From Noisy Labels
Jiangfan Han, Ping Luo, Xiaogang Wang
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 |
|---|---|---|---|
| add self-learning with multi-prototype label correction to cross entropy training | ResNet50 · Clothing1M (1M noisy) | accuracy (%): 69.54 -> 74.45 | Improved |
| add self-learning label correction | ResNet50 · Food101-N | accuracy (%): 84.51 -> 85.11 | Improved |
| add self-learning label correction with 25k verification labels for prototype selection | ResNet50 · Clothing1M (1M noisy + 25k verify) | accuracy (%): 74.69 -> 76.44 | Improved |
| add label correction when 50k clean labels are also used | ResNet50 · Clothing1M (1M noisy + 50k clean) | accuracy (%): 80.27 -> 81.16 | Improved |
| use multiple class prototypes (two instead of one) (ablation) | SMP · Clothing1M | accuracy (%): +2.04% | Improved |
| combine original noisy label with corrected label (alpha=0.5) instead of noisy only or corrected only (ablation) | SMP · Clothing1M | accuracy (%) | Improved |
| use only corrected labels (alpha=1) instead of joint (ablation) | SMP · Clothing1M | accuracy (%) | Hurt |
| increase prototype-selection sample count m from 320 to 2560 (ablation) | SMP · Clothing1M (1M noisy) | accuracy (%): 74.37 -> 74.27 | No clear effect |
| use proposed density/cosine prototype selection instead of K-means++ (ablation) | SMP · Clothing1M (1M noisy) | accuracy (%): 74.08 -> 74.45 | Improved |
| use cosine similarity instead of Euclidean distance density peak (ablation) | SMP · Clothing1M (1M noisy + 25k verify) | accuracy (%): 76.05 -> 76.44 | Improved |
| run label correction iteratively through training (ablation) | SMP · Clothing1M label accuracy | label accuracy (%): 74.38 -> 77.36 | Improved |
Closest papers by meaning
- Coresets for Robust Training of Deep Neural Networks against Noisy Labels (NeurIPS 2020 · similarity 0.91)
- Distilling Effective Supervision From Severe Label Noise (CVPR 2020 · similarity 0.90)
- Toward Robustness against Label Noise in Training Deep Discriminative Neural Networks (NeurIPS 2017 · similarity 0.90)
- Joint Optimization Framework for Learning With Noisy Labels (CVPR 2018 · similarity 0.89)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.87)
- Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources (ICML 2024 · 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.
