Research map / Architectures and efficiency / Neural architecture search
MobileNetV4: Universal Models for the Mobile Ecosystem
Danfeng Qin, Chas H Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, Vaibhav Aggarwal, Tenghui Zhu and 2 more
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 Multi-Head Self-Attention with Multi-Query Attention in the last stage (ablation) | MNv4-Conv-L · ImageNet-1K val | Top-1 (%): 85.27 -> 85.24 | No clear effect |
| replace MHSA with MQA in the last stage (ablation) | MNv4-Conv-L · Pixel 7 EdgeTPU | attention latency (ms): 9.69 ms -> 5.16 ms | Improved |
| add stride-2 spatial down-sampling on keys and values in Mobile MQA (ablation) | MNv4-Hybrid-M · Samsung S23 GPU | latency (ms): 7.4 -> 5.9 | Improved |
| search with full UIB instead of Inverted Bottleneck only (ablation) | NAS · ImageNet-1K val | Top-1 (%): 82.3 -> 83.3 | Improved |
| search with full UIB instead of ConvNext-Like only (ablation) | NAS · ImageNet-1K val | Top-1 (%): 83.2 -> 83.3 | No clear effect |
| use two-stage (coarse then fine) TuNAS search instead of one-stage (ablation) | TuNAS · ImageNet-1K val | Top-1 val (%): 81.26 -> 81.48 | Improved |
| train NAS SuperNet on JFT distillation data instead of ImageNet-1k (ablation) | TuNAS · ImageNet-1K val | MACs (G): 7.2G -> 6.2G | Improved |
| mix D1 and D2 distillation datasets instead of D2 alone (ablation) | MNv4-Conv-L distillation · ImageNet-1K val | Top-1 (%): 84.1 -> 84.4 | Improved |
| add class-balanced JFT data D3 to the ImageNet mixture (ablation) | MNv4-Conv-L distillation · ImageNet-1K val | Top-1 (%): 84.4 -> 84.9 | Improved |
| train distillation student for 2000 epochs instead of 400 (ablation) | MNv4-Conv-L distillation · ImageNet-1K val | Top-1 (%): 84.9 -> 85.9 | Improved |
| use the enhanced distillation recipe instead of the previous SOTA distillation | MNv4-Hybrid-L · ImageNet-1K val | Top-1 (%): 85.7 -> 86.6 | Improved |
| use MNv4-Hybrid-M instead of MNv4-Conv-M as RetinaNet backbone | MNv4 · COCO val | AP: 32.6 -> 34.0 | Improved |
Closest papers by meaning
- MnasNet: Platform-Aware Neural Architecture Search for Mobile (CVPR 2019 · similarity 0.90)
- MobileOne: An Improved One Millisecond Mobile Backbone (CVPR 2023 · similarity 0.87)
- NPAS: A Compiler-Aware Framework of Unified Network Pruning and Architecture Search for Beyond Real-Time Mobile Acceleration (CVPR 2021 · similarity 0.87)
- FreeNet: Liberating Depth-Wise Separable Operations for Building Faster Mobile Vision Architectures (AAAI 2025 · similarity 0.83)
- SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training (CVPR 2025 · similarity 0.83)
- Multi-objective Differentiable Neural Architecture Search (ICLR 2025 · similarity 0.83)
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.
