Research map / Architectures and efficiency / Mixture of experts
Parameters as Experts: Adapting Vision Models with Dynamic Parameter Routing
Meng Lou, Stanley Yu, Yizhou Yu
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 |
|---|---|---|---|
| share one expert center across all 18 layers of a stage instead of groups of 3 layers (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 46.3 -> 47.3 | Improved |
| share expert center across 9 layers instead of 18 (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 47.1 -> 47.3 | Improved |
| activate all experts instead of sparse top-K (ablation) | ParaX · COCO2017 detection, Swin-B | APb: K=3 46.3 -> all 47.4 | Improved |
| use softmax instead of ReLU in the router (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 45.8 -> 47.3 | Improved |
| use softmax instead of Sigmoid in the router (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 46.2 -> 47.3 | Improved |
| use multi-scale kernels [3,5,7] instead of a single kernel (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 46.3 -> 47.1 | Improved |
| add spatially-varying aggregation (SA) module (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 47.1 -> 47.3 | Improved |
| increase kernel sizes to [5,7,9] (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 47.1 -> 46.9 | No clear effect |
| add residual connections to sequential multi-kernel convs (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 46.6 -> 47.3 | Improved |
| change weight init (kaiming normal / uniform vs trunc normal) (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 47.3 vs 47.1 / 47.2 | No clear effect |
| trade latent dim for expert capacity (M=L, C=128 vs M=4L, C=40) (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 46.1 -> 47.3 | Improved |
| increase router hidden size from 16 to 40 (ablation) | ParaX · COCO2017 detection, Swin-B | APb: 47.3 -> 47.4 | No clear effect |
Closest papers by meaning
- Time- Memory- and Parameter-Efficient Visual Adaptation (CVPR 2024 · similarity 0.87)
- pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation (ICLR 2025 · similarity 0.87)
- VFM-Adapter: Adapting Visual Foundation Models for Dense Prediction with Dynamic Hybrid Operation Mapping (AAAI 2025 · similarity 0.86)
- Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation (NeurIPS 2024 · similarity 0.86)
- Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation (CVPR 2025 · similarity 0.86)
- VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024 · similarity 0.85)
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.
