Research map / Architectures and efficiency / Mixture of experts
Scaling Vision with Sparse Mixture of Experts
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace dense MLPs with sparse MoE layers (V-MoE-H/14 Every-2 vs ViT-H/14) | ViT-H/14 · JFT-300M | precision@1: 56.68 -> 60.62 | Helped |
| replace dense MLPs with sparse MoE layers | ViT-H/14 · ImageNet 5-shot | accuracy: 76.95 -> 78.21 | Helped |
| replace dense MLPs with sparse MoE layers | ViT-H/14 · ImageNet fine-tuning | accuracy: 88.08 -> 88.36 | Helped |
| V-MoE-L/16 vs ViT-H/14 with about half the pre-training compute | ViT · ImageNet 5-shot | accuracy: 76.95 -> 77.10 | Helped |
| V-MoE-L/16 vs ViT-H/14 with about half the pre-training compute | ViT · ImageNet fine-tuning | accuracy: 88.08 -> 87.41 | Hurt |
| apply top-k after softmax instead of before (ablation) | V-MoE routing (ViT-S/32, k=2) · ImageNet 5-shot | accuracy: 50.21 -> 53.11 | Helped |
| apply top-k after softmax instead of before (ablation) | V-MoE routing (ViT-S/32, k=2) · JFT-300M | precision@1: 33.75 -> 34.15 | Helped |
| use fewer MoE layers (Last-n instead of Every-2) (ablation) | V-MoE · JFT / ImageNet | quality | No clear effect |
| use Batch Prioritized Routing instead of vanilla routing at low capacity | V-MoE-H/14 · ImageNet | accuracy | Helped |
| apply BPR with max-weight routing during training at reduced capacity | V-MoE · training FLOPs | accuracy vs FLOPs: around 20% of training FLOPs saved | Helped |
| scale to 15B parameters with JFT-3B | V-MoE-15B · ImageNet fine-tuning | accuracy: 90.35 vs MPL 90.20 | Helped |
Closest papers by meaning
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.92)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.91)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (CVPR 2025 · similarity 0.90)
- Mixture of Tokens: Continuous MoE through Cross-Example Aggregation (NeurIPS 2024 · similarity 0.89)
- MoE Jetpack: From Dense Checkpoints to Adaptive Mixture of Experts for Vision Tasks (NeurIPS 2024 · similarity 0.88)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.88)
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.