Research map / Architectures and efficiency / Pruning and efficient tuning
Group Fisher Pruning for Practical Network Compression
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 |
|---|---|---|---|
| prune coupled channels together rather than pruning isolated channels only (ablation) | ResNet-50 · ImageNet validation | inference time (ms): 50.22 -> 36.50 | Helped |
| prune coupled channels together rather than pruning isolated channels only (ablation) | RetinaNet · COCO validation | inference time (ms): 28.49 -> 25.36 | Helped |
| normalize channel importance by memory reduction instead of FLOPs reduction (ablation) | ResNet-50 · ImageNet validation | inference time (ms): 47.07 -> 36.50 | Helped |
| normalize channel importance by memory reduction instead of FLOPs reduction (ablation) | MobileNetV2 · ImageNet validation | inference time (ms): 37.58 -> 24.72 | Helped |
| normalize channel importance by memory reduction instead of FLOPs reduction (ablation) | RetinaNet · COCO validation | inference time (ms): 30.08 -> 25.36 | Helped |
| normalize importance by memory reduction rather than using raw Fisher scores (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy (%): 74.93 -> 76.42 | Helped |
| normalize importance by memory reduction rather than using raw Fisher scores (ablation) | MobileNetV2 · ImageNet validation | top-1 accuracy (%): 72.94 -> 75.97 | Helped |
| use Fisher information instead of absolute first-order gradient as the importance metric (ablation) | ResNet-50 · ImageNet validation | top-1 accuracy (%): 75.8 -> 76.4 | Helped |
| prune ResNet-50 globally with Group Fisher Pruning | ResNet-50 · ImageNet validation | top-1 accuracy (%): 75.06 -> 76.42 | Helped |
| prune MobileNetV2 with Group Fisher Pruning rather than uniform width scaling | MobileNetV2 · ImageNet validation | top-1 accuracy (%): 72.92 -> 73.42 | Helped |
| prune Faster R-CNN with Group Fisher Pruning rather than Slimmable Networks | Faster R-CNN · COCO validation | AP (%): 36.1 -> 37.8 | Helped |
Closest papers by meaning
- DFPC: Data flow driven pruning of coupled channels without data. (ICLR 2023 · similarity 0.89)
- Group Sparsity: The Hinge Between Filter Pruning and Decomposition for Network Compression (CVPR 2020 · similarity 0.88)
- Network Pruning via Performance Maximization (CVPR 2021 · similarity 0.88)
- Dynamic Structure Pruning for Compressing CNNs (AAAI 2023 · similarity 0.88)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.86)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 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.