Research map / Label-efficient and robust learning / Noisy labels
Me-Momentum: Extracting Hard Confident Examples From Noisily Labeled Data
Yingbin Bai, 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 |
|---|---|---|---|
| re-initialize the classifier in an outer loop while keeping extracted confident examples (ablation) | Me-Momentum · CIFAR10 and CIFAR100 noisy labels | label precision of confident examples: label precision increases significantly after re-initialization | Improved |
| add inner-loop alternating refinement of classifier and confident examples (ablation) | Me-Momentum · CIFAR10, CIFAR100 | number of confident examples and test accuracy: accuracy and number increase, label precision decreases slightly | Mixed |
| use the proposed slope-based early stopping instead of highest noisy validation accuracy for initialization (ablation) | Me-Momentum step 1 · CIFAR10 | stopping epoch / noise fitted: stops earlier and fits less noise | Improved |
| vary early-stopping threshold tau over 0.05, 0.07, 0.1, 0.3, 0.5 (ablation) | Me-Momentum · CIFAR10 | accuracy: robust to tau in 0.05 to 0.5 | No clear effect |
| use noisy validation data instead of clean validation data (ablation) | Me-Momentum (pre-trained) · Clothing1M | accuracy: 75.18% -> 73.13% | Hurt |
| pre-train ResNet-50 on ImageNet instead of random initialization (ablation) | Me-Momentum · Clothing1M (clean val) | accuracy: 74.75% -> 75.18% | Improved |
| replace T-revision with Me-Momentum | T-revision · Clothing1M | accuracy: 74.18% -> 75.18% | Improved |
| replace CDR with Me-Momentum | CDR · CIFAR100 Sym-40% | accuracy: 60.18% -> 63.48% | Improved |
| replace T-revision with Me-Momentum | T-revision · CIFAR10 Inst-40% | accuracy: 85.37% -> 86.66% | Improved |
| replace DMI with Me-Momentum | DMI · MNIST Sym-20% | accuracy: 98.92% -> 98.94% | No clear effect |
| replace DMI with Me-Momentum | DMI · MNIST Inst-40% | accuracy: 97.58% -> 98.11% | Improved |
Later papers that compared with this method
4 later accepted papers report Me-Momentum as a baseline in their results tables; 4 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)
- Late Stopping: Avoiding Confidently Learning from Mislabeled Examples (ICCV 2023 · better on 8 of 9 shared results)
- CS-Isolate: Extracting Hard Confident Examples by Content and Style Isolation (NeurIPS 2023 · better on 5 of 5 shared results)
- Self-Filtering: A Noise-Aware Sample Selection for Label Noise with Confidence Penalization (ECCV 2022 · better on 7 of 7 shared results)
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
- Characterizing Datapoints via Second-Split Forgetting (NeurIPS 2022 · similarity 0.86)
- Subclass-Dominant Label Noise: A Counterexample for the Success of Early Stopping (NeurIPS 2023 · similarity 0.86)
- Dataset Distillation for Memorized Data: Soft Labels can Leak Held-Out Teacher Knowledge (ICLR 2026 · similarity 0.86)
- Can Neural Network Memorization Be Localized? (ICML 2023 · similarity 0.85)
- Towards Memorization Estimation: Fast, Formal and Free (ICML 2025 · similarity 0.84)
- Rethinking Learning from Label Proportions via Moment Matching (NeurIPS 2026 · similarity 0.83)
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.
