Research map / Architectures and efficiency / CNN design
Designing Extremely Memory-Efficient CNNs for On-device Vision and Audio Tasks
Yoel Park, Jaewook Lee, Seulki Lee
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 central patch tunnel (ablation) | Proposed network · ImageNet-1k (fp32) | top-1 accuracy (%): 62.32 -> 63.84 | Improved |
| add central patch tunnel (ablation) | Proposed network · VWW | accuracy (%): 78.97 -> 82.80 | Improved |
| use input segmentation into overlapping patches instead of resized whole image (ablation) | Proposed network · ImageNet-1k (fp32) | top-1 accuracy (%): 60.70 -> 63.84 | Improved |
| increase patch margin from 0 to 18 (ablation) | Proposed network · ImageNet-1k (fp32) | top-1 accuracy (%): 61.31 -> 63.84 | Improved |
| increase patch margin from 0 to 9 (ablation) | Proposed network · VWW (fp32) | accuracy (%): 80.00 -> 82.80 | Improved |
| increase expansion ratio t from 2 to 8 with bottleneck reordering (ablation) | Proposed network · ImageNet-1k (fp32) | top-1 accuracy (%): 58.82 -> 63.84 | Improved |
| apply int8 quantization instead of fp32 | Proposed network · ImageNet-1k | top-1 accuracy (%): 63.84 -> 61.58 | Hurt |
| increase expansion ratio t from 2 to 8 (ablation) | Proposed network · ImageNet-1k (int8) | top-1 accuracy (%): 55.75 -> 61.58 | Improved |
Closest papers by meaning
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.89)
- Memory-Optimal Direct Convolutions for Maximizing Classification Accuracy in Embedded Applications (ICML 2019 · similarity 0.88)
- MnasNet: Platform-Aware Neural Architecture Search for Mobile (CVPR 2019 · similarity 0.88)
- MicroNet: Improving Image Recognition With Extremely Low FLOPs (ICCV 2021 · similarity 0.88)
- Memory-efficient Patch-based Inference for Tiny Deep Learning (NeurIPS 2021 · similarity 0.88)
- LSNet: See Large, Focus Small (CVPR 2025 · 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.
