Research map / Label-efficient and robust learning / Noisy labels
Late Stopping: Avoiding Confidently Learning from Mislabeled Examples
Suqin Yuan, Lei Feng, Tongliang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| select mislabeled examples by FkL instead of large loss (ablation) | Late Stopping · CIFAR-100 Ins. 40% (40k-45k range) | label precision: 42.56% -> 91.76% | Improved |
| select mislabeled examples by FkL instead of loss (ablation) | Late Stopping · CIFAR-10 Sym. 40% (40k-45k) | label precision: 91.54% -> 99.94% | Improved |
| select clean examples by FkL instead of loss (ablation) | Late Stopping · CIFAR-100 Sym. 40% (0-10k) | label precision: 88.43% -> 97.45% | Improved |
| late stopping with FkL vs Me-Momentum for retaining clean examples | final training set · CIFAR-100 Sym. 40% | clean examples retained: 22,244 -> 26,120 | Improved |
| late stopping vs Me-Momentum on CIFAR-10 clean retention | final training set · CIFAR-10 Sym. 40% | clean examples retained: 25,791 -> 26,562 | No clear effect |
| Late Stopping vs Me-Momentum on easy CIFAR-10 Sym. 20% | classifier · CIFAR-10 Sym. 20% | test accuracy: 91.44 -> 91.06 | Hurt |
| Late Stopping vs best baselines on CIFAR-100 Ins. 40% | classifier · CIFAR-100 Ins. 40% | test accuracy: 58.38 -> 59.28 | Improved |
| decrease iteration rate m (10 to 2) (ablation) | Late Stopping · CIFAR-10 40% Sym./Ins. | test accuracy | Improved |
| use Late Stopping as pre-processing to cut noise 40% to 20% | training set · CIFAR-100 Ins. | clean examples retained: 27249 -> 25079 (-7.96%) | Mixed |
| Late Stopping prolonged training on larger set | training cost · CIFAR-100 Sym. 40% | training hours: 7.1h -> 9.5h vs Me-Momentum | Hurt |
Later papers that compared with this method
1 later accepted paper reports Late Stopping as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples (NeurIPS 2025 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Me-Momentum — Me-Momentum: Extracting Hard Confident Examples From Noisily Labeled Data (ICCV 2021) · better on 8 of 9 shared results
- Co-teaching — Co-teaching: Robust training of deep neural networks with extremely noisy labels (NeurIPS 2018) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Subclass-Dominant Label Noise: A Counterexample for the Success of Early Stopping (NeurIPS 2023 · similarity 0.90)
- Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples (NeurIPS 2025 · similarity 0.89)
- Understanding and Improving Early Stopping for Learning with Noisy Labels (NeurIPS 2021 · similarity 0.89)
- Revisiting Learning with Noisy Labels: Active Forgetting and Noise Suppression (CVPR 2026 · similarity 0.89)
- Early Stopping Against Label Noise Without Validation Data (ICLR 2024 · similarity 0.88)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · 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.
