Research map / Label-efficient and robust learning / Noisy labels
Understanding and Mitigating Human-Labelling Errors in Supervised Contrastive Learning
Zijun Long, Lipeng Zhuang, George W Killick, Richard Mccreadie, Gerardo Aragon-Camarasa, Paul Henderson
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 |
|---|---|---|---|
| replace SupCon loss with SCL-RHE (down-weight easy positives, correct for mislabels) | BEiT-3 from scratch · ImageNet-1K | top-1 accuracy: 82.57 -> 84.21 | Improved |
| replace SupCon with SCL-RHE | BEiT-3 from scratch · CIFAR-100 | top-1 accuracy: 60.77 -> 64.47 | Improved |
| use synthetic-noise methods Sel-CL / TCL instead of SupCon on low-noise human-labelled data | BEiT-3 from scratch · ImageNet-1K | top-1 accuracy: 82.57 -> 81.87 (Sel-CL), 81.74 (TCL) | Hurt |
| replace cross-entropy with supervised contrastive objectives on small data | BEiT-3 from scratch · CIFAR-10 | top-1 accuracy: 71.70 -> 90.16 | Improved |
| evaluate on corrected test labels instead of original | SCL-RHE · ImageNet-1K | top-1 accuracy: 84.21 -> 86.02 (+1.81) | Improved |
| Sel-CL and TCL evaluated on corrected test labels | Sel-CL/TCL · ImageNet-1K | top-1 accuracy: 81.87 -> 81.35 (Sel-CL), 81.74 -> 81.28 (TCL) | Hurt |
| fine-tune pre-trained BEiT-3 with SCL-RHE instead of SupCon | BEiT-3 · iNat2017 | top-1 accuracy: 72.85 -> 76.25 | Improved |
| fine-tune BEiT-3 with Sel-CL or TCL instead of cross-entropy | BEiT-3 · CUB-200 | top-1 accuracy: 98.00 -> 94.52 (Sel-CL), 93.89 (TCL) | Hurt |
| use SCL-RHE instead of Sel-CL for training time | ResNet-18 contrastive training · CIFAR (time per epoch) | min/epoch: 29.1 -> 20.4 | Improved |
| apply SCL-RHE at 20% synthetic label noise | ResNet-18 · CIFAR-10 (20% noise) | top-1 accuracy: 94.45 (Sel-CL) -> 94.71 | Improved |
| apply SCL-RHE at extreme 40% synthetic noise instead of Sel-CL/TCL | ResNet-18 · CIFAR-100 (40% noise) | top-1 accuracy: 74.62 (TCL) -> 74.08 | Hurt |
| also handle false negatives in negative set (mitigation for negatives) | SCL-RHE · label-error analysis | share of incorrect learning signals: 99.04% from false positives on CIFAR-100; 99.91% on ImageNet-1K | No clear effect |
Closest papers by meaning
- Human-Corrected Labels Learning: Enhancing Labels Quality via Human Correction of VLMs Discrepancies (AAAI 2026 · similarity 0.89)
- FFF: Fixing Flawed Foundations in Contrastive Pre-Training Results in Very Strong Vision-Language Models (CVPR 2024 · similarity 0.86)
- How does Labeling Error Impact Contrastive Learning? A Perspective from Data Dimensionality Reduction (ICML 2025 · similarity 0.86)
- Revisiting Interpolation for Noisy Label Correction (AAAI 2025 · similarity 0.86)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.86)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · 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.
