Research map / Label-efficient and robust learning / Noisy labels
Unsupervised Label Noise Modeling and Loss Correction
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 dynamic hard bootstrapping instead of static hard bootstrapping (ablation) | PreAct ResNet-18 · CIFAR-10, 80% label noise | last validation accuracy: 25.7 to 64.2 | Helped |
| combine mixup with dynamic hard bootstrapping and regularization | PreAct ResNet-18 · CIFAR-10, 80% label noise | best validation accuracy: 71.6 to 86.8 | Helped |
| combine mixup with dynamic hard bootstrapping and regularization | PreAct ResNet-18 · CIFAR-100, 80% label noise | best validation accuracy: 30.8 to 48.2 | Helped |
| add dynamic input mixup guided by clean probabilities (ablation) | M-DYR-H · CIFAR-10, 90% label noise | best validation accuracy: 40.8 to 56.7 | Helped |
| progressively lower softmax temperature from soft to hard bootstrapping (ablation) | MD-DYR-H · CIFAR-10, 90% label noise | best validation accuracy: 56.7 to 69.1 | Helped |
| progressively lower softmax temperature from soft to hard bootstrapping (ablation) | MD-DYR-H · CIFAR-100, 90% label noise | best validation accuracy: 13.5 to 24.3 | Helped |
| replace BMM-estimated bootstrapping weights with fixed weights (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 74.6 | Hurt |
| replace the beta mixture model with a Gaussian mixture model (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 83.5 | Hurt |
| re-estimate BMM every 5 epochs instead of every epoch (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 86.8 | No clear effect |
| re-estimate BMM every 0.5 epochs instead of every epoch (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 87.5 | No clear effect |
| increase EM iterations from 10 to 5 (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 87.2 | No clear effect |
| increase EM iterations from 10 to 20 (ablation) | M-DYR-H · CIFAR-10, 80% label noise | last validation accuracy: 86.6 to 86.3 | No clear effect |
Closest papers by meaning
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.89)
- Sample Selection with Uncertainty of Losses for Learning with Noisy Labels (ICLR 2022 · similarity 0.88)
- L2B: Learning to Bootstrap Robust Models for Combating Label Noise (CVPR 2024 · similarity 0.88)
- Mitigating Label Noise through Data Ambiguation (AAAI 2024 · similarity 0.87)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.86)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 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.