Research map / Architectures and efficiency / CNN design
Filter Grafting for Deep Neural Networks
Fanxu Meng, Hao Cheng, Ke Li, Zhixin Xu, Rongrong Ji, Xing Sun, Guangming Lu
Check 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 |
|---|---|---|---|
| Graft external network weights into MobileNetV2 | MobileNetV2 · CIFAR-100 | accuracy: 71.44 -> 72.90 | Improved |
| Use external scions rather than internal filters (ablation) | MobileNetV2 filter grafting · CIFAR-10 | accuracy: 92.68 -> 92.94 | Improved |
| Use entropy rather than L1 norm to measure layer information (ablation) | ResNet32 grafting · CIFAR-10 | accuracy: 93.24 -> 93.33 | Improved |
| Use entropy rather than L1 norm to measure layer information (ablation) | MobileNetV2 grafting · CIFAR-100 | accuracy: 72.90 -> 73.26 | Improved |
| Vary sampling order across grafted models (ablation) | MobileNetV2 grafting · CIFAR-100 | accuracy: 71.91 -> 73.26 | Improved |
| Vary learning rates as well as sampling order (ablation) | MobileNetV2 grafting · CIFAR-100 | accuracy: 73.26 -> 74.15 | Improved |
| Increase the number of grafted networks from 6 to 8 (ablation) | MobileNetV2 grafting · CIFAR-100 | accuracy: 78.32 -> 77.76 | Hurt |
| Graft complete layer weights rather than selected filters (ablation) | ResNet32 grafting · CIFAR-100 | accuracy: 70.79 -> 71.28 | Improved |
| Calculate entropy over the whole layer rather than summing filter entropies (ablation) | ResNet32 grafting · CIFAR-100 | accuracy: 70.95 -> 71.28 | Improved |
| Use adaptive weighting rather than a fixed coefficient of 0.5 (ablation) | ResNet32 grafting · CIFAR-100 | accuracy: 70.70 -> 71.28 | Improved |
| Graft weights using six networks rather than one baseline network | MobileNetV2 · CIFAR-100 | accuracy: 71.44 -> 78.32 | Improved |
| Graft weights between networks | ResNet18 · ImageNet | top-1: 69.15 -> 71.19 | Improved |
Closest papers by meaning
- Learning Versatile Filters for Efficient Convolutional Neural Networks (NeurIPS 2018 · similarity 0.86)
- Provable Filter Pruning for Efficient Neural Networks (ICLR 2020 · similarity 0.86)
- ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression (ICCV 2017 · similarity 0.86)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.86)
- RePr: Improved Training of Convolutional Filters (CVPR 2019 · similarity 0.85)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.85)
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.
