AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Collaboration of Experts: Achieving 80% Top-1 Accuracy on ImageNet with 100M FLOPs

Yikang Zhang, Zhuo Chen, Zhao Zhong

ICML 2022 · Architectures and efficiency › Mixture of experts · approach: Sparse MoE as regularization and attention · 7 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add CoE expert collaboration to OFA-230OFA-230 · ImageNet validationTop-1 accuracy: 78.0% to 80.7%Improved
Add CondConv to CoE-LargeCoE-Large · ImageNet validationTop-1 accuracy: 80.7% -> 81.5%Improved
Replace gate-value routing with CoE's label-supervised trainingOFA-230 with four experts · ImageNet validationTop-1 accuracy: 78.7% -> 79.9%; 78.9% -> 79.9%Improved
Replace category-based routing with learned CoE selectionOFA-230 with four experts · ImageNet validationTop-1 accuracy: 78.3% -> 79.9%; 77.5% -> 79.9%Improved
Add model-level CoE instead of averaging four model predictionsOFA-230 · ImageNet validationTop-1 accuracy: 79.6% -> 80.7%Improved
Add LGM for expert selection (ablation)CoE-Large · ImageNet validationTop-1 accuracy: 78.0% -> 79.9%Improved
Add WGM to assign weighted training portions to experts (ablation)CoE-Large · ImageNet validationTop-1 accuracy: 78.1% -> 79.9%Improved
Partition training data by delegator output rather than expert suitability (ablation)CoE-Large WGM · ImageNet validationTop-1 accuracy: 77.0% -> 79.9%Improved
Add suitability standardization and equal-expert assignment constraint (ablation)CoE-Large LGM · ImageNet validationTop-1 accuracy: 79.4% -> 79.9%Improved
Add equal assignment and smoothing in WGM (ablation)CoE-Large WGM · ImageNet validationTop-1 accuracy: 79.2% -> 79.9%Improved
Use progressive rather than constant assignment sharpening (ablation)CoE-Large WGM · ImageNet validationTop-1 accuracy: 79.4% -> 79.9%Improved
Reweight expert-selection loss by sample suitability differences (ablation)CoE-Large · ImageNet validationTop-1 accuracy: 79.5% -> 79.9%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy