Research map / Label-efficient and robust learning / Noisy labels
Learning from Training Dynamics: Identifying Mislabeled Data beyond Manually Designed Features
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 the learned training-dynamics detector to identify and exclude suspected mislabeled samples | ResNet-32 trained on CIFAR-10 with symmetric label noise · CIFAR-10, symmetric noise ratios 0.2, 0.4, 0.6, 0.8 | test accuracy: 90.2$\pm$0.0 -> 91.1$\pm$0.0; 87.5$\pm$0.1 -> 88.9$\pm$0.0; 82.1$\pm$0.0 -> 83.5$\pm$0.2; 54.4$\pm$1.6 -> 55.5$\pm$1.4 | Helped |
| Use the learned training-dynamics detector to identify and exclude suspected mislabeled samples | ResNet-32 trained on CIFAR-100 with symmetric label noise · CIFAR-100, symmetric noise ratios 0.2, 0.4, 0.6, 0.8 | test accuracy: 65.5$\pm$0.2 -> 65.6$\pm$0.3; 61.3$\pm$0.1 -> 60.8$\pm$0.2; 53.0$\pm$0.5 -> 54.3$\pm$0.4; 31.7$\pm$0.7 -> 30.1$\pm0.6$ | Mixed |
| Use the learned training-dynamics detector to identify and exclude suspected mislabeled samples | ResNet models on public noisy datasets · CIFAR-10, CIFAR-100, Tiny ImageNet, WebVision50, Clothing100K | test accuracy: 92.1$\pm$0.0 -> 92.5$\pm$0.1; 68.2$\pm$0.1 -> 68.7$\pm$0.3; 51.4$\pm$0.1 -> 52.2$\pm$0.0; 80.2 -> 81.4; 66.5 -> 68.2 | Helped |
| Fine-tune the pretrained noise detector on the target dataset | CUB-200-2011 and Caltech-256 noise detection · Synthesized noisy CUB-200-2011 and Caltech-256 | mAP | Helped |
| Add 5% data debugging to DM-AugDesc | DM-AugDesc · CUB-200-2011 20% symmetric noise | Best accuracy: 79.65 -> 79.55 | Hurt |
| Add 10% data debugging to DM-AugDesc | DM-AugDesc · CUB-200-2011 20% symmetric noise | Best accuracy: 79.65 -> 80.14 | Helped |
| Add data debugging to DM-AugDesc | DM-AugDesc · mini WebVision | Top1 accuracy: 78.64 -> 79.32; 78.64 -> 78.84 | Helped |
| Add data debugging to DM-AugDesc | DM-AugDesc · mini WebVision | Top5 accuracy: 93.20 -> 92.56; 93.20 -> 92.84 | Hurt |
| Add 5% data debugging to DivideMix | DivideMix · CUB-200-2011 20% symmetric noise | Best accuracy: 75.96 -> 76.06 | Helped |
| Add 10% data debugging to DivideMix | DivideMix · CUB-200-2011 20% symmetric noise | Best accuracy: 75.96 -> 76.77 | Helped |
| Add data debugging to DivideMix | DivideMix · mini WebVision | Top1 accuracy: 77.32 -> 77.84; 77.32 -> 78.24 | Helped |
| Add data debugging to DivideMix | DivideMix · mini WebVision | Top5 accuracy: 91.64 -> 92.16; 91.64 -> 92.72 | Helped |
Closest papers by meaning
- Learning Discriminative Dynamics with Label Corruption for Noisy Label Detection (CVPR 2024 · similarity 0.88)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.87)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.87)
- IOFM: Using the Interpolation Technique on the Over-Fitted Models to Identify Clean-Annotated Samples (AAAI 2024 · similarity 0.87)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.87)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (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.