Research map / Label-efficient and robust learning / Recognition and long-tail
MDCS: More Diverse Experts with Consistency Self-distillation for Long-tailed Recognition
Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, Jun 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 |
|---|---|---|---|
| add Diversity Loss (diversity softmax per expert) to multi-expert training (ablation) | MDCS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 47.8 -> 50.7 | Improved |
| add RandAug weak-strong augmentation on top of DL (ablation) | MDCS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 50.7 -> 54.1 | Improved |
| add Consistency Self-distillation on top of DL and RandAug (ablation) | MDCS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 54.1 -> 56.1 | Improved |
| add CS and RandAug without Diversity Loss (ablation) | MDCS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 47.8 -> 48.5 | No clear effect |
| use weak-strong views for CS instead of strong-strong (ablation) | CS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 54.6 -> 56.1 | Improved |
| use weak-strong views for CS instead of weak-weak (ablation) | CS · CIFAR100-LT (IF 100) | top-1 accuracy (%): 50.7 -> 56.1 | Improved |
| add CS to reduce expert model variance (ablation) | MDCS · CIFAR100-LT (IF 100) | model variance (all): 0.40 -> 0.36 | Improved |
| add CS to improve ensemble diversity and expert accuracy (ablation) | MDCS · CIFAR100-LT (IF 100) | diversity factor sigma: 53.4 -> 60.4 | Improved |
| use diverse lambda {-0.5,1,2.5} instead of all zeros (ablation) | Diversity Loss · CIFAR100-LT (IF 100) | diversity factor all-class (%): 55.4 -> 66.9 | Improved |
| use lambda {2,2,2} (strongly reversed weights) (ablation) | Diversity Loss · CIFAR100-LT (IF 100) | diversity factor all-class (%): 66.9 -> 58.2 | Hurt |
| increase number of experts (ablation) | MDCS · CIFAR100-LT (IF 100) | accuracy | Improved |
| set CS loss weight alpha to 0.6 (ablation) | MDCS · CIFAR100-LT (IF 100) | accuracy | Improved |
Later papers that compared with this method
3 later accepted papers report MDCS as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Trust-calibrated Collaborative Learning for Long-Tailed Visual Recognition (CVPR 2026 · better on 5 of 6 shared results)
- VICAL: Vicinal Consistency Alignment for Long-Tailed Visual Recognition (ECCV 2026 · better on 1 of 1 shared result)
- Echoes of the Past: Boosting Long-tail Recognition via Reflective Learning (ECCV 2024 · better on 2 of 2 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
- Self-Supervised Aggregation of Diverse Experts for Test-Agnostic Long-Tailed Recognition (NeurIPS 2022 · similarity 0.89)
- Self Supervision to Distillation for Long-Tailed Visual Recognition (ICCV 2021 · similarity 0.89)
- Harnessing Hierarchical Label Distribution Variations in Test Agnostic Long-tail Recognition (ICML 2024 · similarity 0.88)
- Long-tailed Recognition with Model Rebalancing (NeurIPS 2025 · similarity 0.88)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.88)
- LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content (CVPR 2024 · 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.
