Research map / Architectures and efficiency / Mixture of experts
Towards Understanding the Mixture-of-Experts Layer in Deep Learning
Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, Yuanzhi Li
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 |
|---|---|---|---|
| replace single nonlinear expert with nonlinear MoE (8 experts) | CNN with cubic activation · Synthetic setting 1 | test accuracy: 79.48 -> 99.46 | Improved |
| replace single linear expert with linear MoE | linear CNN · Synthetic setting 1 | test accuracy: 68.71 -> 92.99 | Improved |
| use nonlinear instead of linear experts in MoE (ablation) | MoE · Synthetic setting 1 | test accuracy: 92.99 -> 99.46 | Improved |
| use nonlinear instead of linear experts in MoE (stronger noise) (ablation) | MoE · Synthetic setting 2 | test accuracy: 88.48 -> 98.09 | Improved |
| use nonlinear experts so router dispatch entropy drops (ablation) | MoE · Synthetic setting 1 | dispatch entropy: 1.300 -> 0.098 | Improved |
| add load balancing loss to linear MoE (ablation) | linear MoE · Synthetic setting 1 | test accuracy: 92.99 -> 93.81 | No clear effect |
| MoE vs single model on CIFAR-10 (no cluster structure) | ResNet18 · CIFAR-10 | accuracy: 95.51 -> 95.32 | No clear effect |
| MoE vs single model on CIFAR-10-Rotate | ResNet18 · CIFAR-10-Rotate | accuracy: 88.23 -> 92.60 | Improved |
| MoE vs single model on CIFAR-10-Rotate | CNN · CIFAR-10-Rotate | accuracy: 76.78 -> 79.60 | Improved |
| MoE vs single classifier on multilingual sentiment text embeddings | MoE classifier · multilingual sentiment | accuracy: 74.13 -> 76.22 | Improved |
Closest papers by meaning
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.91)
- Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning (ICML 2025 · similarity 0.90)
- MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts (AAAI 2024 · similarity 0.90)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (CVPR 2025 · similarity 0.89)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.89)
- Tight Clusters Make Specialized Experts (ICLR 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.
