Research map / Detection and segmentation / CNN features
Slimmable Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train MobileNet v1 jointly across widths using switchable batch normalization | MobileNet v1 · ImageNet validation, 1.0× | top-1 error: 29.1 -> 28.5 | Helped |
| Train MobileNet v2 jointly across widths using switchable batch normalization | MobileNet v2 · ImageNet validation, 1.0× | top-1 error: 28.2 -> 29.5 | Hurt |
| Train ShuffleNet jointly across widths using switchable batch normalization | ShuffleNet · ImageNet validation, 2.0× | top-1 error: 26.3 -> 28.7 | Hurt |
| Train ResNet-50 jointly across widths using switchable batch normalization | ResNet-50 · ImageNet validation, 1.0× | top-1 error: 23.9 -> 24.0 | Hurt |
| Use a slimmable backbone instead of individually trained backbones | Faster R-CNN · COCO 2017 validation, 1.0× | Box AP: 36.4 -> 36.8 | Helped |
| Use a slimmable backbone instead of individually trained backbones | Mask R-CNN · COCO 2017 validation, 1.0× | Mask AP: 34.2 -> 34.9 | Helped |
| Use a slimmable backbone instead of individually trained backbones | Keypoints R-CNN · COCO 2017 validation, 1.0× | Kps AP: 61.3 -> 63.9 | Helped |
| Share BN statistics across widths rather than using switchable batch normalization (ablation) | Slimmable networks · ImageNet, 1000-class classification | top-1 accuracy: around 0.1% | Hurt |
| Train wider MobileNet v2 progressively while fixing the base model weights (ablation) | MobileNet v2 · ImageNet validation, 0.35× to 0.5× | top-1 accuracy: 60.3% -> 61.0%; individual 0.5×: 65.4% | Hurt |
| Use private scale and bias for each width instead of shared scale and bias (ablation) | S-MobileNet v1 · ImageNet validation, 0.25× | top-1 error: 47.1 -> 46.9 | Helped |
| Use private scale and bias for each width instead of shared scale and bias (ablation) | S-MobileNet v1 · ImageNet validation, 0.5× | top-1 error: 35.9 -> 35.2 | Helped |
| Use private scale and bias for each width instead of shared scale and bias (ablation) | S-MobileNet v1 · ImageNet validation, 0.75× | top-1 error: 30.9 -> 30.5 | Helped |
Closest papers by meaning
- Dynamic Slimmable Network (CVPR 2021 · similarity 0.87)
- Three Guidelines You Should Know for Universally Slimmable Self-Supervised Learning (CVPR 2023 · similarity 0.86)
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.85)
- Slimmable Generative Adversarial Networks (AAAI 2021 · similarity 0.84)
- DeepRebirth: Accelerating Deep Neural Network Execution on Mobile Devices (AAAI 2018 · similarity 0.84)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.82)
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.