Research map / Label-efficient and robust learning / Noisy labels
Learning to Learn From Noisy Labeled Data
Check 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 |
|---|---|---|---|
| apply MLNT training rather than reproduced cross-entropy training | CIFAR-10 · CIFAR-10 test, symmetric and asymmetric label noise | accuracy (%) | Helped |
| use MLNT-teacher after three training iterations rather than the strongest baseline | MLNT · CIFAR-10 test, symmetric noise r=0.9 | accuracy (%): 58.0 -> 59.09±1.12 | Helped |
| use MLNT-teacher after three training iterations rather than the strongest baseline | MLNT · CIFAR-10 test, asymmetric noise r=0.5 | accuracy (%): 84.6 -> 82.09±0.47 | Hurt |
| use MLNT-teacher after three training iterations rather than Cross Entropy | MLNT · Clothing1M test | accuracy (%): +4.19 | Helped |
| use MLNT-teacher after three training iterations rather than Joint Optimization | MLNT · Clothing1M test | accuracy (%): +1.31 | Helped |
| add iterative training using the previous iteration's model for filtering and refined consistency targets (ablation) | MLNT · CIFAR-10 and Clothing1M | accuracy (%) | Helped |
| increase the number of synthetic noisy mini-batches M (ablation) | MLNT-student (1st iter.) · CIFAR-10, symmetric and asymmetric noise | accuracy (%) | Helped |
| increase M from 10 to 15 (ablation) | MLNT-student (1st iter.) · CIFAR-10, symmetric and asymmetric noise | accuracy (%) | No clear effect |
| vary the fraction of synthetic labels replaced, ρ (ablation) | MLNT-student (1st iter.) · CIFAR-10, asymmetric label noise | accuracy (%) | No clear effect |
| use first-order approximation instead of full second-order optimization (ablation) | MLNT-student (1st iter.) · CIFAR-10, four representative noise settings | accuracy (%): 90.57 vs 90.74; 81.96 vs 82.05; 91.84 vs 91.89; 88.70 vs 88.91 | No clear effect |
| continue iterative training beyond three iterations (ablation) | MLNT · CIFAR-10 and Clothing1M | accuracy (%) | No clear effect |
| add the meta-objective to conventional training on clean CIFAR-10 (ablation) | MLNT · CIFAR-10 test, clean training data | accuracy (%): +1.68 | Helped |
Closest papers by meaning
- Training Noise-Robust Deep Neural Networks via Meta-Learning (CVPR 2020 · similarity 0.91)
- Learning to Purify Noisy Labels via Meta Soft Label Corrector (AAAI 2021 · similarity 0.90)
- Joint Optimization Framework for Learning With Noisy Labels (CVPR 2018 · similarity 0.90)
- Augmentation Strategies for Learning With Noisy Labels (CVPR 2021 · similarity 0.89)
- Meta Label Correction for Noisy Label Learning (AAAI 2021 · similarity 0.89)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (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.