Research map / Label-efficient and robust learning / Semi-supervised learning
DivideMix: Learning with Noisy Labels as Semi-supervised Learning
Junnan Li, Richard Socher, Steven C.H. Hoi
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 |
|---|---|---|---|
| Use DivideMix instead of M-correction | CIFAR-10 with symmetric label noise · CIFAR-10, symmetric noise at 20%, 50%, 80%, and 90% | Best test accuracy (%): 94.0 -> 96.1; 92.0 -> 94.6; 86.8 -> 93.2; 69.1 -> 76.0 | Improved |
| Use DivideMix instead of M-correction at 20% and 50%, and Meta-Learning at 80% and 90% | CIFAR-100 with symmetric label noise · CIFAR-100, symmetric noise at 20%, 50%, 80%, and 90% | Best test accuracy (%): 73.9 -> 77.3; 66.1 -> 74.6; 42.4 -> 60.2; 24.3 -> 31.5 | Improved |
| Use DivideMix instead of Meta-Learning | CIFAR-10 with asymmetric label noise · CIFAR-10, asymmetric noise 40% | Best test accuracy (%): 89.2 -> 93.4 | Improved |
| Use DivideMix instead of P-correction | Clothing1M · Clothing1M | Test accuracy (%): 73.49 -> 74.76 | Improved |
| Use DivideMix instead of Iterative-CV | WebVision and ILSVRC12 validation · WebVision and ILSVRC12 validation | Top-1 accuracy (%): 65.24 -> 77.32; 61.60 -> 75.20 | Improved |
| Use an ensemble of both networks at test time instead of only network θ(1) (ablation) | DivideMix · CIFAR-10 and CIFAR-100 label-noise benchmarks | Test accuracy (%) | Improved |
| Use co-training instead of self-divide (ablation) | DivideMix · CIFAR-100, symmetric noise 90% | Best test accuracy (%): 27.7 -> 31.5 | Improved |
| Add label refinement (ablation) | DivideMix · CIFAR-10, asymmetric noise 40% | Best test accuracy (%): 87.7 -> 93.4 | Improved |
| Add input augmentation (ablation) | DivideMix · CIFAR-10, asymmetric noise 40% | Best test accuracy (%): 89.5 -> 93.4 | Improved |
| Replace co-divide with self-divide and use original MixMatch (ablation) | DivideMix · CIFAR-100, symmetric noise 90% | Best test accuracy (%): 25.0 -> 31.5 | Hurt |
Later papers that compared with this method
68 later accepted papers report DivideMix as a baseline in their results tables; 65 of them report a better number on most of the results they share with it.
- Deconstructing the Failure of Ideal Noise Correction: A Three-Pillar Diagnosis (CVPR 2026 · better on 2 of 6 shared results)
- Content Diversity-guided Ambiguity Mitigation for Open-Set Noisy Label Learning (AAAI 2026 · better on 6 of 6 shared results)
- Enhancing Learning with Noisy Labels via Rockafellian Relaxation (ICLR 2026 · better on 2 of 2 shared results)
- See Through the Noise: Improving Domain Generalization in Gaze Estimation (CVPR 2026 · better on 2 of 2 shared results)
- Jump-teaching: Combating Sample Selection Bias via Temporal Disagreement (AAAI 2026 · better on 2 of 2 shared results)
- Variation-Bounded Loss for Noise-Tolerant Learning (AAAI 2026 · better on 2 of 2 shared results)
- TrainRef: Curating Data with Label Distribution and Minimal Reference for Accurate Prediction and Reliable Confidence (ICLR 2026 · better on 1 of 1 shared result)
- TANGO: Text-Anchored Guided Optimization for Robust Fine-tuning Vision-Language Models under Label Noise (CVPR 2026 · better on 1 of 1 shared result)
60 more not listed.
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
- SplitNet: Learnable Clean-Noisy Label Splitting for Learning with Noisy Labels (IJCV 2024 · similarity 0.89)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.86)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.86)
- RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · similarity 0.86)
- Label-Efficient Dataset Pruning via Semi-Supervised Pseudo-Labeling (NeurIPS 2026 · similarity 0.85)
- Directional Label Diffusion Model for Learning from Noisy Labels (CVPR 2025 · similarity 0.85)
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.
