Research map / Architectures and efficiency / Mixture of experts
Efficient Deweahter Mixture-of-Experts with Uncertainty-Aware Feature-Wise Linear Modulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoFME, a parameter-efficient mixture-of-experts for multi-weather image restoration that replaces parallel FFN experts with feature-modulation experts on one shared FFN, plus an MC-dropout uncertainty-aware router.
Problem
Conventional MoE with parallel FFN experts is too parameter- and compute-heavy for practical deployment in multi-task adverse weather removal, and naive linear routers poorly calibrate task-specific feature assignment.
Approach
Implicitly instantiate experts as learnable feature-wise affine modulations (FM) on a single shared FFN (FME), with routing weights calibrated by MC-dropout uncertainty (UaR) to assign tokens to modulation modules.
Key results
- MoFME reaches 29.01 dB average PSNR and 0.9243 SSIM on All-Weather, outperforming baselines by at least 0.2 dB per the paper's text (MoWE 28.95 dB).
- On RainCityscapes MoFME gets 31.34 dB average PSNR with 17.68M parameters versus MoWE's 31.19 dB at 44.84M parameters.
- MoFME saves more than 74% of parameters and about 20% inference time relative to MoWE-style MoE, per the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoFME instead of MoWE | multi-task deweathering · RainCityscapes average | PSNR: 31.19 -> 31.34 | Helped |
| Use MoFME instead of MoWE | multi-task deweathering · All-Weather average | PSNR: 28.95 -> 29.01 | Helped |
| Replace parallel FFN experts with FME and add UaR (ablation) | MoWE · All-Weather average | PSNR: 28.95 -> 29.01 | Helped |
| Add UaR while omitting FME (ablation) | MoWE · All-Weather average | PSNR: 28.95 -> 29.18 | Helped |
| Replace parallel FFN experts with FME while omitting UaR (ablation) | MoWE · All-Weather average | PSNR: 28.95 -> 28.68 | Hurt |
| Replace parallel FFN experts with FME and add UaR (ablation) | MoE · All-Weather average | PSNR: 28.36 -> 28.49 | Helped |
| Replace parallel FFN experts with FME and add UaR (ablation) | MMoE · All-Weather average | PSNR: 28.50 -> 28.66 | Helped |
| Use MoFME instead of efficient MoE baselines | downstream Cityscapes segmentation · Cityscapes derain | mIoU | Helped |
| Use MoFME instead of similar-size baselines | Swin-T. image classification · Tiny-ImageNet | top-1 accuracy: 0.8% | Helped |
Surprising findings
- Table 2 reports MoWE as best on RainCityscapes Derain PSNR (32.99 vs MoFME 32.52), and M3ViT underperforms on deweather despite being designed for dense tasks.
- The ablation shows FME alone lowers average PSNR for the MoWE base (28.68 vs 29.19 with UaR only), so the parameter savings come with a performance trade-off in some settings.
Limitations
- Experiments use only 8 experts in most settings; scaling to hundreds of experts is claimed but not evaluated.
- Efficiency gains are reported only for the deweathering setting with ViT backbones; generalization to other domains is limited to classification and segmentation downstream.
Experimental setup
Datasets: All-Weather RainCityscapes Cityscapes Tiny-ImageNet CIFAR-10
Benchmarks: All-Weather RainCityscapes Cityscapes Tiny-ImageNet CIFAR-10
Metrics: PSNR SSIM mIoU mAcc top-1 accuracy parameters FLOPs inference time
Baselines: MoE-ViT MMoE-ViT M3ViT MoWE OneS PR-MoE Restormer TransWeather PReNet MSBDN-DFF
Builds on: Mixture-of-Experts feature-wise linear modulation M3ViT MoWE MC dropout
Closest papers by meaning
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.86)
- From Sparse to Soft Mixtures of Experts (ICLR 2024 · similarity 0.86)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.86)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.86)
- Multi-Head Mixture-of-Experts (NeurIPS 2024 · similarity 0.85)
- Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 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.