Research map / Architectures and efficiency / Attention and transformers
Rethinking Mobile Block for Efficient Attention-based Models
Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, Chengjie Wang
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 |
|---|---|---|---|
| add DW-Conv to iRMB (EW-MHSA only vs both) (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: 76.6 -> 78.4 | Improved |
| add EW-MHSA to iRMB (DW-Conv only vs both) (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: 77.6 -> 78.4 | Improved |
| replace Transformer blocks with iRMB | DeiT-Tiny · ImageNet-1K val | Top-1 accuracy: 72.2 -> 74.3 | Improved |
| replace Transformer blocks with iRMB | PVT-Tiny · ImageNet-1K val | Top-1 accuracy: 75.1 -> 75.4 | Improved |
| use one-residual iRMB instead of two-residual MetaFormer with iRMB token mixer (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: 77.5 -> 78.4 | Improved |
| cascade DW-Conv and EW-MHSA instead of parallel (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: 78.1 -> 78.4 | Improved |
| put EW-MHSA before DW-Conv rather than switching order (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: -0.6 when order switched | Improved |
| use stronger training recipes (DeiT, EdgeNeXt) instead of simple recipe (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy: 78.4 vs 78.1 / 78.3 | No clear effect |
| apply MHSA in all four stages instead of last two (ablation) | EMO-5M · ImageNet-1K val | Top-1 accuracy / FLOPs: 78.4 -> 78.8, FLOPs 903M -> 992M | Mixed |
| use LN in all stages instead of stages 3/4 (ablation) | EMO-5M · ImageNet-1K val | throughput: 1,693 -> 952 throughput, +0.2 Top-1 | Hurt |
| use EMO block instead of EdgeNeXt block for mobile inference | EMO-5M · iPhone14 latency (ms) | latency: 27.7 -> 6.8 | Improved |
| replace operators with group conv, asymmetric conv or performer (ablation) | iRMB · ImageNet-1K val | Top-1 accuracy | No clear effect |
Closest papers by meaning
- EMOv2: Pushing 5M Vision Model Frontier (TPAMI 2025 · similarity 0.93)
- Neural Architecture Search for Lightweight Non-Local Networks (CVPR 2020 · similarity 0.86)
- MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models (ICLR 2023 · similarity 0.86)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.85)
- Mobile Attention: Mobile-Friendly Linear-Attention for Vision Transformers (ICML 2024 · similarity 0.85)
- MobileMamba: Lightweight Multi-Receptive Visual Mamba Network (CVPR 2025 · similarity 0.84)
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.
