Research map / Label-efficient and robust learning / Noisy labels
Mitigating Memorization of Noisy Labels by Clipping the Model Prediction
Hongxin Wei, HUIPING ZHUANG, RENCHUNZI XIE, Lei Feng, Gang Niu, Bo An, Yixuan Li
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 |
|---|---|---|---|
| add LogitClip to CE loss | CE · CIFAR-10 instance-dependent noise | test accuracy: 68.26 to 86.74 | Improved |
| add LogitClip to CE loss | CE · CIFAR-10 symmetric-50% and other noise types | test accuracy: e.g. 70.88 to 84.37 (Sym-50%) | Improved |
| add LogitClip to CE loss | CE · CIFAR-100 noise settings | test accuracy: e.g. 47.07 to 63.16 (Sym-50%) | Improved |
| add LogitClip to robust losses (MAE, PHuber-CE, SCE, GCE, Taylor-CE, NCE, AEL, AUL, Cores, NCE+MAE, NCE+AGCE) | robust losses · CIFAR-10 and CIFAR-100 | test accuracy: large gains on many settings (e.g. AUL Sym-20% 76.73 to 91.31) but near-zero or slight drops on some (SCE Sym-20% 91.48 | Mixed |
| add LogitClip to CE loss | CE on WebVision Mini ResNet-18 · WebVision Mini | top-1 validation accuracy (best epoch): 62.6 to 65.12 | Improved |
| add LogitClip to DivideMix | DivideMix · CIFAR-10 noisy settings | test accuracy: 94.41 to 95.15 (Sym-50%) | Improved |
| clip logit norm with tau small (0.25–1) instead of large tau on clean CIFAR-100 (ablation) | CE + LogitClip · clean CIFAR-100 | accuracy: 75.98 (CE) to 46.04 at tau=0.25, 68.00 at tau=0.5 | Hurt |
| use a large tau (20) for LogitClip on clean CIFAR-100 (ablation) | CE + LogitClip · clean CIFAR-100 | accuracy: 75.90 vs CE 75.98 | No clear effect |
| replace norm clipping with clipping-by-value on logits (ablation) | CE · noisy CIFAR-10 (instance-dependent, real) | test accuracy | Hurt |
| replace LogitClip with LogitNorm (exact norm constraint) (ablation) | CE · CIFAR-10 asymmetric noise | test accuracy: 81.81 (LogitNorm) vs 86.91 (LogitClip) | Hurt |
| add Norm Regularization (L2 logit penalty) instead of LogitClip (ablation) | CE · noisy CIFAR-10 across noise types | test accuracy | Hurt |
| apply ReLU6-style clamping of intermediate activations (ablation) | CE · noisy CIFAR-10 | test accuracy | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DivideMix — DivideMix: Learning with Noisy Labels as Semi-supervised Learning (ICLR 2020) · better on 1 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
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.90)
- Learning from Noisy Labels with Complementary Loss Functions (AAAI 2021 · similarity 0.89)
- Can gradient clipping mitigate label noise? (ICLR 2020 · similarity 0.89)
- Learning With Noisy Labels via Sparse Regularization (ICCV 2021 · similarity 0.88)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.87)
- Handling Label Noise via Instance-Level Difficulty Modeling and Dynamic Optimization (NeurIPS 2025 · 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.
