Research map / Label-efficient and robust learning / Noisy labels
Tackling Instance-Dependent Label Noise via a Universal Probabilistic Model
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 |
|---|---|---|---|
| model instance-dependent label noise with learnable confusing probabilities and alternating label-posterior updates | ResNet-32 · CIFAR-100 synthetic IDN | test accuracy (%): 46.02$\pm$0.42 -> 47.51$\pm$0.28 | Helped |
| model instance-dependent label noise with learnable confusing probabilities and alternating label-posterior updates | ResNet-32 · CIFAR-10 synthetic IDN | test accuracy (%): 69.51$\pm$0.63 -> 69.82$\pm$0.20 | Helped |
| use the confusing-instance probabilistic model rather than the strongest listed baseline | ResNet-32 · CIFAR-10 synthetic CCN, r=0.1 | test accuracy (%): 93.27$\pm$0.10 -> 93.81$\pm$0.05 | Helped |
| use the confusing-instance probabilistic model rather than the strongest listed baseline | ResNet-32 · CIFAR-10 synthetic CCN, r=0.2 | test accuracy (%): 92.86$\pm$0.15 -> 93.06$\pm$0.10 | Helped |
| use the confusing-instance probabilistic model rather than the strongest listed baseline | ResNet-32 · CIFAR-10 synthetic CCN, r=0.3 | test accuracy (%): 91.64$\pm$0.11 -> 92.79$\pm$0.21 | Helped |
| use the confusing-instance probabilistic model rather than the strongest listed baseline | ResNet-32 · CIFAR-10 synthetic CCN, r=0.4 | test accuracy (%): 90.39$\pm$0.16 -> 90.48$\pm$0.16 | Helped |
| use the confusing-instance probabilistic model instead of Forward | ResNet-32 · CIFAR-10 synthetic CCN, r=0.5 | test accuracy (%): 80.22$\pm$0.39 -> 78.03$\pm$0.55 | Hurt |
| use the confusing-instance probabilistic model rather than PENCIL | ResNet-50 · Clothing1M | test accuracy (%): 73.49 -> 74.02 | Helped |
| add 50k correctly labeled training instances, fixing their confusing probabilities to zero (ablation) | Ours · Clothing1M | test accuracy (%): 74.02 -> 77.55 | Helped |
| finetune the noisy-data-trained classifier using 50k correctly labeled instances (ablation) | Ours · Clothing1M | test accuracy (%): 74.02 -> 80.68 | Helped |
| use a confusing-probability learning rate of 0.5 instead of 50 (ablation) | Ours · CIFAR-100 validation | validation accuracy | Helped |
| use a very small confusing-probability learning rate such as 0.01 (ablation) | Ours · CIFAR-100 validation | validation accuracy | Hurt |
Closest papers by meaning
- Learning with Instance-Dependent Label Noise: A Sample Sieve Approach (ICLR 2021 · similarity 0.93)
- A Second-Order Approach to Learning With Instance-Dependent Label Noise (CVPR 2021 · similarity 0.93)
- Noise-Robust Learning from Multiple Unsupervised Sources of Inferred Labels (AAAI 2022 · similarity 0.93)
- Understanding Instance-Level Label Noise: Disparate Impacts and Treatments (ICML 2021 · similarity 0.92)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · similarity 0.90)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.89)
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.