Research map / Label-efficient and robust learning / Noisy labels
Clusterability as an Alternative to Anchor Points When Learning with 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 |
|---|---|---|---|
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Global versus strongest listed baseline · Inst. CIFAR-10, η=0.2 | clean test accuracy (%): 90.31$\pm$0.37 -> 89.71$\pm$0.51 | Hurt |
| Estimate local rather than one global transition matrix | HOC · Inst. CIFAR-10, η=0.2/0.4/0.6 | clean test accuracy (%): 89.71$\pm$0.51 -> 90.03$\pm$0.15; 84.62$\pm$1.02 -> 85.49$\pm$0.80; 70.67$\pm$3.38 -> 77.40$\pm$0.47 | Helped |
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Global versus strongest listed baseline · Inst. CIFAR-10, η=0.4 | clean test accuracy (%): 84.99$\pm$0.81 -> 84.62$\pm$1.02 | Helped |
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Local versus strongest listed baseline · Inst. CIFAR-10, η=0.6 | clean test accuracy (%): 75.15$\pm$0.82 -> 77.40$\pm$0.47 | Helped |
| Use a global rather than local transition matrix | HOC · Inst. CIFAR-100, η=0.2/0.4/0.6 | clean test accuracy (%): 67.47$\pm$0.85 -> 68.82$\pm$0.26; 61.20$\pm$1.04 -> 62.29$\pm$1.11; 49.84$\pm$1.81 -> 52.96$\pm$1.85 | Helped |
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Global versus strongest listed baseline · Inst. CIFAR-100, η=0.2 | clean test accuracy (%): 58.00$\pm$0.20 -> 68.82$\pm$0.26 | Helped |
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Global versus strongest listed baseline · Inst. CIFAR-100, η=0.4 | clean test accuracy (%): 48.01$\pm$0.12 -> 62.29$\pm$1.11 | Helped |
| Estimate the transition matrix using 2-NN noisy-label consensuses and apply forward loss correction | HOC Global versus strongest listed baseline · Inst. CIFAR-100, η=0.6 | clean test accuracy (%): 40.88$\pm$7.57 -> 52.96$\pm$1.85 | Helped |
| Estimate the transition matrix with HOC and use forward loss correction | HOC · Clothing1M | test accuracy (%): 73.24 -> 73.39 | Helped |
| Estimate the transition matrix with HOC and use forward loss correction | HOC · Human CIFAR-10 | test accuracy (%): 90.39 -> 90.62 | Helped |
| Use a standard CE-trained representation extractor under instance-dependent noise | 2-NN clusterability tuples · CIFAR-10 and CIFAR-100, η=0.2/0.4 | feasible 2-NN tuple ratio (%): CIFAR-10: 87.88, 89.06, 78.15, 79.85; CIFAR-100: 82.82, 84.33, 64.88, 68.31 | Mixed |
Closest papers by meaning
- Beyond Images: Label Noise Transition Matrix Estimation for Tasks with Lower-Quality Features (ICML 2022 · similarity 0.90)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.88)
- Learning the Latent Causal Structure for Modeling Label Noise (NeurIPS 2024 · similarity 0.88)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.88)
- Noisy Label Learning with Instance-Dependent Outliers: Identifiability via Crowd Wisdom (NeurIPS 2024 · similarity 0.87)
- Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label Configurations (NeurIPS 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.