Research map / Label-efficient and robust learning / Noisy labels
Curriculum Loss: Robust Learning and Generalization against Label Corruption
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 NPCL instead of GCE, pruning high-loss samples during training | single-network training · MNIST, Symmetry-50% | test accuracy: 92.48% ± 0.13% -> 98.53% ± 0.02% | Helped |
| Use NPCL instead of GCE, pruning high-loss samples during training | single-network training · CIFAR10, Symmetry-50% | test accuracy: 61.80% ± 0.11% -> 77.66% ± 0.09% | Helped |
| Use NPCL instead of GCE | single-network training · CIFAR10, Symmetry-20% | test accuracy: 84.68% ± 0.05% -> 84.30% ± 0.07% | Hurt |
| Use NPCL instead of Co-teaching+ | single-network training · MNIST, Pair-35% | test accuracy: 93.81% ± 0.20% -> 97.90% ± 0.04% | Helped |
| Use NPCL instead of Co-teaching+ | single-network training · CIFAR100, Symmetry-20% | test accuracy: 56.15% ± 0.09% -> 55.30% ± 0.09% | Hurt |
| Replace NPCL with CL (do not prune based on a noise-rate prior) | DenseNet · CIFAR10, uniform (40%) noise | test accuracy: 83.24 -> 85.57 | Helped |
| Replace NPCL with CL | DenseNet · CIFAR100, uniform (60%) noise | test accuracy: 44.49 -> 46.65 | Helped |
| Replace CL with NPCL | DenseNet · CIFAR100, semantic noise, DenseNet label generator (34%) | test accuracy: 55.14 -> 57.59 | Helped |
| Set the NPCL noise-rate prior to 1.25 times the estimated rate instead of the true rate (ablation) | NPCL · MNIST, Symmetry-20% | test accuracy: 99.41% ± 0.01% -> 99.55% ± 0.02% | Helped |
| Set the NPCL noise-rate prior to 0.5 times the estimated rate instead of the true rate (ablation) | NPCL · MNIST, Symmetry-50% | test accuracy: 98.53% ± 0.02% -> 78.67% ± 0.36% | Hurt |
| Set the NPCL noise-rate prior to 1.5 times the estimated rate instead of the true rate (ablation) | NPCL · MNIST, Symmetry-50% | test accuracy: 98.53% ± 0.02% -> 67.61% ± 0.06% | Hurt |
| Replace the hard multi-class hinge loss with the proposed soft multi-class hinge loss (ablation) | CIFAR100 training · CIFAR100 | test accuracy | Helped |
Closest papers by meaning
- Robust Curriculum Learning: from clean label detection to noisy label self-correction (ICLR 2021 · similarity 0.89)
- SuperLoss: A Generic Loss for Robust Curriculum Learning (NeurIPS 2020 · similarity 0.88)
- Toward Understanding Adversarial Distillation: Why Robust Teachers Fail (ICML 2026 · similarity 0.87)
- How does Disagreement Help Generalization against Label Corruption? (ICML 2019 · similarity 0.86)
- Robust Classification via Regression for Learning with Noisy Labels (ICLR 2024 · similarity 0.86)
- Curriculum Learning by Transfer Learning: Theory and Experiments with Deep Networks (ICML 2018 · 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.