Research map / Architectures and efficiency / Mixture of experts
Teacher-Guided Routing for Sparse Vision Mixture-of-Experts
Masahiro Kada, Ryota Yoshihashi, Satoshi Ikehata, Rei Kawakami, Ikuro Sato
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-guided routing | VMoE Tiny · ImageNet-1K | top-1 accuracy: 77.85 -> 78.78 | Improved |
| add teacher-guided routing | VMoE Small · ImageNet-1K | top-1 accuracy: 82.63 -> 83.34 | Improved |
| add teacher-guided routing | VMoE Base · ImageNet-1K | top-1 accuracy: 83.97 -> 85.46 | Improved |
| add z-loss regularization | VMoE Tiny · ImageNet-1K | top-1 accuracy: 77.85 -> 77.99 | No clear effect |
| keep TGR-MoE during fine-tuning (not only pre-training) (ablation) | TGR-MoE Tiny · CIFAR-100 | top-1 accuracy: 86.15 -> 86.95 | Improved |
| keep TGR-MoE during fine-tuning (not only pre-training) (ablation) | TGR-MoE Small · CIFAR-100 | top-1 accuracy: 89.18 -> 90.26 | Improved |
| keep TGR-MoE during fine-tuning (not only pre-training) (ablation) | TGR-MoE Base · CIFAR-100 | top-1 accuracy: 90.05 -> 91.07 | Improved |
| add teacher-guided routing at 64 experts | VMoE Tiny · ImageNet-1K (64 experts) | top-1 accuracy: 78.74% -> 79.95% | Improved |
| add teacher-guided routing at 128 experts | VMoE Tiny · ImageNet-1K (128 experts) | top-1 accuracy: 79.38% -> 80.36% | Improved |
| apply distillation only in first half of training instead of throughout (ablation) | TGR-MoE Tiny 8 experts · ImageNet-1K | top-1 accuracy: 77.81 -> 78.13 | Improved |
| remove task loss from router training (distillation only) (ablation) | TGR-MoE Tiny 8 experts · ImageNet-1K | top-1 accuracy: 77.81 -> 77.83 | No clear effect |
| use teacher final-layer features instead of intermediate features for routing signal (ablation) | TGR-MoE Tiny 8 experts · ImageNet-1K | top-1 accuracy: 77.81 -> 75.83 | Hurt |
Closest papers by meaning
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.91)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.90)
- Tight Clusters Make Specialized Experts (ICLR 2025 · similarity 0.88)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (CVPR 2025 · similarity 0.88)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.88)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 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.
