Research map / Label-efficient and robust learning / Knowledge distillation
Deep Ensemble Learning by Diverse Knowledge Distillation for Fine-Grained Object Classification
Naoki Okamoto, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi
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 |
|---|---|---|---|
| optimized diverse-KD graph instead of independent training (2 nodes, ABN) | ensemble of ABN · Stanford Dogs | ensemble accuracy (%): 70.90 -> 73.86 | Improved |
| optimized diverse-KD graph instead of DML (4 nodes, ABN) | ensemble of ABN · Stanford Dogs | ensemble accuracy (%): 72.87 -> 74.16 | Improved |
| optimized graph instead of independent training (5 nodes, ResNet-18) | ensemble of ResNet-18 · Stanford Dogs | ensemble accuracy (%): 69.47 -> 71.35 | Improved |
| DML instead of independent training (ResNet-18, 5 nodes) | ensemble of ResNet-18 · Stanford Dogs | ensemble accuracy (%): 69.47 -> 69.94 | No clear effect |
| ONE multi-branch instead of independent networks (ABN, 5 branches) | ensemble of ABN · Stanford Dogs | ensemble accuracy (%): 72.32 -> 70.89 | Hurt |
| reuse graph optimized on Stanford Dogs for CUB-200-2011 training | 2-node graph · CUB-200-2011 | ensemble accuracy (%): 65.26 -> 72.06 | Improved |
| reuse graph optimized on Stanford Dogs for CIFAR-100 training | 2-node graph · CIFAR-100 | ensemble accuracy (%): 73.16 -> 72.19 | Hurt |
| optimize graph on the same dataset for CIFAR-100 | 2-node graph · CIFAR-100 | ensemble accuracy (%): 73.16 -> 74.18 | Improved |
| graph optimized on CUB-200-2011 itself rather than Stanford Dogs (ablation) | 2-node graph · CUB-200-2011 | ensemble accuracy (%): 72.06 -> 69.81 | Hurt |
| use CIFAR-100 graph on CIFAR-10 instead of Stanford Dogs graph (ablation) | 2-node graph · CIFAR-10 | ensemble accuracy (%): 93.87 -> 94.37 | Improved |
| distill from ensemble of optimized 2-node graph instead of DML teacher | ResNet-18 student · Stanford Dogs | student accuracy (%): 69.68 -> 72.94 | Improved |
| larger divergence between networks' output distributions | ensemble of two ResNets · Stanford Cars | ensemble gain correlation: 0.499 | Improved |
Closest papers by meaning
- Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning (ICLR 2023 · similarity 0.89)
- Knowledge Distillation by On-the-Fly Native Ensemble (NeurIPS 2018 · similarity 0.89)
- Improving Ensemble Distillation With Weight Averaging and Diversifying Perturbation (ICML 2022 · similarity 0.87)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.86)
- From Coarse to Fine: Enable Comprehensive Graph Self-supervised Learning with Multi-granular Semantic Ensemble (ICML 2024 · similarity 0.86)
- Ensemble Distribution Distillation via Flow Matching (ICML 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.
