Research map / Label-efficient and robust learning / Noisy labels
Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels
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 MentorMix instead of the strongest baseline | Inception-ResNet-v2 on Red Mini-ImageNet, fine-tuned · Red Mini-ImageNet, 10 noise levels | peak accuracy (%): 82.7$\pm$1.8 -> 83.3$\pm$1.9 | Helped |
| Combine MentorNet loss-based sampling with Mixup | MentorMix · Blue Mini-ImageNet, 50% noise | peak accuracy (%): +16.3\% | Helped |
| Combine MentorNet loss-based sampling with Mixup | MentorMix · Red Mini-ImageNet, 50% noise | peak accuracy (%): +2.4\% | Helped |
| Use MentorMix rather than the prior state of the art | WebVision full, ILSVRC12 validation · WebVision 1.0 | top-1 accuracy (%): about 3\% | Helped |
| Give the anchor example the larger mixup coefficient when its MentorNet weight is high (ablation) | MentorMix · Noisy-label classification | performance | Helped |
| Use a second MentorNet weighting after mixup (ablation) | MentorMix at high noise levels · Noisy-label classification | performance | Helped |
| Remove the second MentorNet weighting (ablation) | MentorMix at low noise levels · Noisy-label classification | performance | No clear effect |
| Remove image-to-image search results when constructing the benchmark (ablation) | Controlled web-label dataset · Red Mini-ImageNet and Red Stanford Cars | experimental results | No clear effect |
| Train on web-label noise rather than symmetric synthetic noise | Vanilla DNNs · Mini-ImageNet and Stanford Cars | generalization performance | Helped |
| Fine-tune stronger ImageNet-pretrained architectures rather than training from scratch | DNNs on noisy labels · Mini-ImageNet and Stanford Cars | peak accuracy: 0.91; 0.88 | Helped |
| Apply synthetic-noise expectations about early pattern learning to web-label noise | DNNs trained on web-label noise · Stanford Cars | peak-to-final accuracy drop | Mixed |
Closest papers by meaning
- Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotations (ICLR 2022 · similarity 0.91)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.89)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.88)
- Understanding and Mitigating the Label Noise in Pre-training on Downstream Tasks (ICLR 2024 · similarity 0.88)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.88)
- Mitigating Label Noise through Data Ambiguation (AAAI 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.