Research map / Architectures and efficiency / Attention and transformers
An Efficient Hybrid Vision Transformer for TinyML Applications
Fanhong Zeng, Huanan Li, Juntao Guan, Rui Fan, Tong Wu, Xilong Wang, Rui Lai
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 last-stage attention with Squeeze-and-Excitation channel module (ablation) | MobileViT-XXS · ImageNet-1K | top-1 accuracy (%): 70.0 -> 70.8 | Improved |
| replace last-stage attention with SE (ablation) | SHViT-S2 · ImageNet-1K | top-1 accuracy (%): 74.9 -> 75.3 | Improved |
| replace LeanSHSA with 4-head MHSA (ablation) | TinyNeXt-M · ImageNet-1K (100 epochs) | top-1 accuracy (%) and throughput: 74.1 -> 74.3; 4889 -> 3950 img/s | No clear effect |
| replace LeanSHSA with bottlenecked SHSA (ablation) | TinyNeXt-M · ImageNet-1K (100 epochs) | top-1 accuracy (%): 74.1 -> 73.9 | Hurt |
| use spatial attention (MHSA) in the last stage instead of SE (ablation) | TinyNeXt-M · ImageNet-1K (100 epochs) | top-1 accuracy (%): 74.1 -> 74.1; params 2.3M -> 2.7M | No clear effect |
| use SHSA in the last stage instead of SE (ablation) | TinyNeXt-M · ImageNet-1K (100 epochs) | top-1 accuracy (%): 74.1 -> 73.6 | Hurt |
| replace BatchNorm with LayerNorm (ablation) | TinyNeXt-M · ImageNet-1K (100 epochs) | top-1 accuracy (%): 74.1 -> 73.8; throughput 4889 -> 4479 | Hurt |
| use TinyNeXt-S instead of MobileViT-XXS at similar parameters | TinyNeXt-S · ImageNet-1K | top-1 accuracy (%): 69.0 -> 72.7 | Improved |
| use TinyNeXt-M backbone instead of EMO-2M in SSDLite | TinyNeXt-M · COCO 2017 | mAP: 25.2 -> 25.0 | No clear effect |
| use TinyNeXt-S instead of EMO-1M on Jetson Nano | TinyNeXt-S · Jetson Nano latency | latency (ms): 26.2 -> 14.3 | Improved |
Closest papers by meaning
- TinyViT: Fast Pretraining Distillation for Small Vision Transformers (ECCV 2022 · similarity 0.91)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.88)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.87)
- SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · similarity 0.87)
- Real-time Core-Periphery Guided ViT with Smart Data Layout Selection on Mobile Devices (NeurIPS 2024 · similarity 0.87)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.87)
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.
