Research map / Label-efficient and robust learning / Noisy labels
Are Anchor Points Really Indispensable in Label-Noise Learning?
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 |
|---|---|---|---|
| Add T-revision to Reweight (ablation) | Reweight label-noise learning · MNIST, CIFAR-10, and CIFAR-100, Sym-20% and Sym-50% | accuracy: 98.71±0.11 -> 98.91±0.04; 98.13±0.19 -> 98.38±0.21; 86.77±0.40 -> 89.63±0.13; 80.16±0.46 -> 83.40±0.65; 58.35±0.64 -> 65 | Helped |
| Add T-revision to Reweight (ablation) | Reweight label-noise learning without possible anchors · MNIST, CIFAR-10, and CIFAR-100, Sym-20% and Sym-50% | accuracy: 98.69±0.08 -> 98.85±0.02; 98.05±0.22 -> 98.37±0.17; 85.53±0.26 -> 88.90±0.22; 77.70±1.00 -> 81.55±0.94; 56.60±0.71 -> 62 | Helped |
| Add T-revision to Reweight (ablation) | Reweight label-noise learning · Clothing1M | accuracy: 70.95 -> 74.18 | Helped |
| Use risk-consistent Reweight rather than classifier-consistent Forward with revised transition matrix | T-Revision · MNIST, CIFAR-10, CIFAR-100, and Clothing1M | accuracy: 98.84±0.09 -> 98.91±0.04; 98.12±0.22 -> 98.38±0.21; 88.10±0.21 -> 89.63±0.13; 81.11±0.74 -> 83.40±0.65; 62.13±2.09 -> 65 | Mixed |
| Remove high estimated clean-posterior examples, used as possible anchors | Forward and Reweight consistent estimators · CIFAR-100 Sym-20% and Sym-50% | accuracy: Forward-A 57.75±0.37 -> Forward-N/A 56.23±0.34; Reweight-A 58.35±0.64 -> Reweight-N/A 56.60±0.71; Forward-A 44.66±1.01 - | Hurt |
| Remove high estimated clean-posterior examples, used as possible anchors | Decoupling, MentorNet, and Co-teaching · MNIST | accuracy: 95.39±0.29 -> 95.93±0.21; 96.57±0.18 -> 97.11±0.09; 97.22±0.18 -> 97.69±0.23 | Mixed |
| Remove high estimated clean-posterior examples, used as possible anchors | Decoupling, MentorNet, and Co-teaching · CIFAR-10 and CIFAR-100 | accuracy: 79.85±0.30 -> 75.37±1.24; 80.49±0.52 -> 78.51±0.31; 82.38±0.11 -> 81.72±0.14; 42.75±0.49 -> 39.59±0.42; 52.11±0.10 -> 48 | Mixed |
| Increase symmetric label-noise rate | Reweight-A-R · MNIST, Sym-60%, Sym-70%, and Sym-80% | accuracy: 97.83±0.18; 97.13±0.08; 94.19±0.45 | Hurt |
Closest papers by meaning
- Provably End-to-end Label-noise Learning without Anchor Points (ICML 2021 · similarity 0.94)
- Learning the Latent Causal Structure for Modeling Label Noise (NeurIPS 2024 · similarity 0.89)
- Learning Causal Transition Matrix for Instance-dependent Label Noise (AAAI 2025 · similarity 0.89)
- Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification (NeurIPS 2026 · similarity 0.88)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.87)
- Dirichlet-based Per-Sample Weighting by Transition Matrix for Noisy Label Learning (ICLR 2024 · similarity 0.87)
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.