Research map / Label-efficient and robust learning / Noisy labels
SELF: Learning to Filter Noisy Labels with Self-Ensembling
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 progressive filtering to training on the noisy label set (ablation) | ResNet26 · CIFAR-10, 40% symmetric noise | accuracy: 83.20 -> 87.35 | Helped |
| Add Mean Teacher model averaging (ablation) | ResNet26 with progressive filtering · CIFAR-10, 40% symmetric noise | accuracy: 87.35 -> 93.70 | Helped |
| Add ensemble predictions for label filtering (ablation) | Mean Teacher with progressive filtering · CIFAR-10, 80% symmetric noise | accuracy: 57.40 -> 69.91 | Helped |
| Add ensemble predictions for label filtering (ablation) | Mean Teacher with progressive filtering · CIFAR-100, 80% symmetric noise | accuracy: 38,61 -> 42.09 | Helped |
| Continue using filtered samples as unlabeled data instead of removing them completely (ablation) | SELF · CIFAR-10 and CIFAR-100, 80% symmetric noise, noisy validation | accuracy: 59.98 -> 69.91; 35.53 -> 42.09 | Helped |
| Use Mean Teacher rather than the ResNet baseline (ablation) | CIFAR-100 · CIFAR-100, symmetric noise | accuracy: 53.18 -> 65.85; 19.92 -> 26.31 | Mixed |
| Use a clean validation set instead of a noisy validation set | SELF · CIFAR-10 and CIFAR-100, 80% symmetric noise | accuracy: 69.91 -> 79.93; 42.09 -> 46.43 | Helped |
| Use SELF rather than the strongest reported baseline | CIFAR-10 · CIFAR-10, 80% symmetric noise, noisy validation | accuracy: 67.92 -> 69.91 | Helped |
| Use SELF rather than MentorNet | ImageNet-ILSVRC, ResNext50 · ImageNet validation, 40% training noise | P@1: 65.10 -> 71.31 | Helped |
| Use the Mean Teacher loss rather than entropy loss with iterative filtering (ablation) | CIFAR-10 · CIFAR-10, 80% symmetric noise | accuracy: 53.44 -> 69.91 | Helped |
Closest papers by meaning
- Learning with Instance-Dependent Label Noise: A Sample Sieve Approach (ICLR 2021 · similarity 0.89)
- Augmentation Strategies for Learning With Noisy Labels (CVPR 2021 · similarity 0.89)
- A Second-Order Approach to Learning With Instance-Dependent Label Noise (CVPR 2021 · similarity 0.89)
- Tackling Instance-Dependent Label Noise via a Universal Probabilistic Model (AAAI 2021 · similarity 0.89)
- Self-cognitive Denoising in the Presence of Multiple Noisy Label Sources (ICML 2024 · similarity 0.88)
- On the Role of Label Noise in the Feature Learning Process (ICML 2025 · similarity 0.88)
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.