Research map / Label-efficient and robust learning / Noisy labels
CleanNet: Transfer Learning for Scalable Image Classifier Training With 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 |
|---|---|---|---|
| Use CleanNet instead of classification filtering for label-noise detection | Food-101N · Food-101N | average error rate: 16.60 -> 9.61 | Helped |
| Retrain the classifier on data cleansed by CleanNet, then use its features (ablation) | CleanNet label-noise detection · Food-101N | average error rate: 9.61 -> 6.99 | Helped |
| Use soft rather than hard sample weighting (ablation) | CleanNet classifier · Food-101N | top-1 accuracy: 83.47 -> 83.95 | Helped |
| Use CleanNet soft weighting instead of no noise-handling method | ResNet-50 classifier · Food-101N | top-1 accuracy: 81.44 -> 83.95 | Helped |
| Use CleanNet soft weighting rather than no noise-handling method | ResNet-50 classifier · Clothing1M | top-1 accuracy: 68.94 -> 74.69 | Helped |
| Use soft rather than hard sample weighting (ablation) | CleanNet classifier · Clothing1M | top-1 accuracy: 74.15 -> 74.69 | Helped |
| Transfer verification supervision to held-out classes | CleanNet label-noise detection · Food-101N held-out classes | average error rate: 41.5% relative reduction (n=10) | Helped |
| Hold out verification labels in 50 of 101 classes while using transfer learning (ablation) | CleanNet label-noise detection · Food-101N held-out classes | average error rate: 11.02%; 10.12% | Mixed |
| Use CleanNet soft weighting rather than no noise-handling method | ResNet-50 classifier · WebVision validation | top-1 accuracy: 67.76 -> 69.14 | Helped |
| Verify labels for every image rather than using no verification | ResNet-50 classifier · WebVision validation | top-1 accuracy: 67.76 -> 70.31 | Helped |
| Use semantic-308 verification instead of all-1000 verification (ablation) | CleanNet classifier · WebVision validation | top-1 accuracy: 69.14 -> 68.96 | Hurt |
| Use random-308 verification instead of all-1000 verification (ablation) | CleanNet classifier · WebVision validation | top-1 accuracy: 69.14 -> 68.89 | Hurt |
Closest papers by meaning
- Toward Robustness against Label Noise in Training Deep Discriminative Neural Networks (NeurIPS 2017 · similarity 0.90)
- Weakly Supervised Image Classification Through Noise Regularization (CVPR 2019 · similarity 0.89)
- Learning From Noisy Large-Scale Datasets With Minimal Supervision (CVPR 2017 · similarity 0.89)
- SELF: Learning to Filter Noisy Labels with Self-Ensembling (ICLR 2020 · similarity 0.88)
- Delving into Noisy Label Detection with Clean Data (ICML 2023 · similarity 0.88)
- Combating Semantic Contamination in Learning with Label Noise (AAAI 2025 · 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.