Research map / Architectures and efficiency / CNN design
MicroNet: Improving Image Recognition With Extremely Low FLOPs
Yunsheng Li, Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Lu Yuan, Zicheng Liu, Lei Zhang, Nuno Vasconcelos
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 Micro-Factorized depthwise convolution (ablation) | MobileNet-like baseline · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 44.9 -> 46.4 | Improved |
| add Micro-Factorized pointwise convolution (ablation) | MobileNet-like baseline + DW · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 46.4 -> 50.0 | Improved |
| add lite combination at low levels (ablation) | MicroNet static · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 50.0 -> 51.7 | Improved |
| add lite combination at both low and high levels instead of low only (ablation) | MicroNet · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 51.7 -> 51.2 | Hurt |
| add static Shift-Max (ablation) | MicroNet static · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 51.7 -> 54.4 | Improved |
| use dynamic instead of static Shift-Max (ablation) | MicroNet · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 54.4 -> 58.5 | Improved |
| use adaptive group number G=sqrt(C/R) instead of fixed G (ablation) | Micro-Factorized pointwise conv · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 50.8 -> 51.7 | Improved |
| increase group multiplier lambda from 1.0 to 4.0 (ablation) | Micro-Factorized pointwise conv · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 51.7 -> 47.6 | Hurt |
| use dynamic Shift-Max instead of dynamic ReLU (ablation) | MicroNet-M2 · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 56.0 -> 58.5 | Improved |
| use dynamic Shift-Max instead of SE+ReLU (ablation) | MicroNet-M2 · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 54.4 -> 58.5 | Improved |
| add dynamic Shift-Max in all three layers instead of one (A1) (ablation) | MicroNet-M2 · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 55.9 -> 58.5 | Improved |
| increase K from 2 to 3 in dynamic Shift-Max (J=2) (ablation) | MicroNet-M2 · ImageNet (MicroNet-M2 ablation) | top-1 accuracy: 58.5 -> 58.1 | No clear effect |
Closest papers by meaning
- Designing Extremely Memory-Efficient CNNs for On-device Vision and Audio Tasks (IJCV 2026 · similarity 0.88)
- Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification (NeurIPS 2020 · similarity 0.88)
- Is Second-Order Information Helpful for Large-Scale Visual Recognition? (ICCV 2017 · similarity 0.87)
- Memory Matching Networks for One-Shot Image Recognition (CVPR 2018 · similarity 0.87)
- MUXConv: Information Multiplexing in Convolutional Neural Networks (CVPR 2020 · similarity 0.87)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.86)
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.
