Research map / Label-efficient and robust learning / Knowledge distillation
MoE-KD: Your Teacher Model is Worth Mixture-of-Experts for Knowledge Distillation
Kuo Shi, Wenjie Zhu, Bo Peng
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 teacher-classifier gating instead of uniform gating (ablation) | MoE-KD · CIFAR-100 WRN-40-2 -> WRN-40-1 | top-1 accuracy (%): 73.87 -> 75.21 | Improved |
| add teacher-classifier gating instead of uniform gating (ablation) | MoE-KD · CIFAR-100 ResNet50 -> MobileNetV2 | top-1 accuracy (%): 69.28 -> 71.38 | Improved |
| reuse pre-trained teacher classifier as gating instead of learning it from scratch (ablation) | MoE-KD · CIFAR-100 WRN-40-2 -> WRN-40-1 | top-1 accuracy (%): 74.43 -> 75.21 | Improved |
| reuse pre-trained teacher classifier as gating instead of learning it from scratch (ablation) | MoE-KD · CIFAR-100 ResNet50 -> MobileNetV2 | top-1 accuracy (%): 70.62 -> 71.38 | Improved |
| derive expert embeddings from teacher features instead of learning from scratch (ablation) | MoE-KD · CIFAR-100 WRN-40-2 -> WRN-40-1 | top-1 accuracy (%): 74.66 -> 75.21 | Improved |
| derive expert embeddings from teacher features instead of learning from scratch (ablation) | MoE-KD · CIFAR-100 ResNet50 -> MobileNetV2 | top-1 accuracy (%): 70.65 -> 71.38 | Improved |
| use soft aggregation instead of hard aggregation (ablation) | MoE-KD · CIFAR-100 WRN-40-2 -> WRN-40-1 | top-1 accuracy (%): 74.95 -> 75.21 | Improved |
| use soft aggregation instead of hard aggregation (ablation) | MoE-KD · CIFAR-100 ResNet50 -> MobileNetV2 | top-1 accuracy (%): 71.16 -> 71.38 | Improved |
| use Bayes-optimal EM E-step instead of direct teacher assignment (ablation) | MoE-KD · CIFAR-100 WRN-40-2 -> WRN-40-1 | top-1 accuracy (%): 74.79 -> 75.21 | Improved |
| use Bayes-optimal EM E-step instead of direct teacher assignment (ablation) | MoE-KD · CIFAR-100 ResNet50 -> MobileNetV2 | top-1 accuracy (%): 70.94 -> 71.38 | Improved |
| use MoE-KD instead of DiffKD | MoE-KD · ImageNet-1K ResNet34 -> ResNet18 | top-1 accuracy (%): 72.22 -> 73.15 | Improved |
| use MoE-KD instead of DiffKD | MoE-KD · ImageNet-1K ResNet50 -> MobileNetV1 | top-1 accuracy (%): 73.62 -> 73.41 | Hurt |
Closest papers by meaning
- Toward Student-oriented Teacher Network Training for Knowledge Distillation (ICLR 2024 · similarity 0.90)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · similarity 0.90)
- Knowledge Distillation with Auxiliary Variable (ICML 2024 · similarity 0.89)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.89)
- Can Students Beyond the Teacher? Distilling Knowledge from Teacher’s Bias (AAAI 2025 · similarity 0.89)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.89)
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.
