Research map / Architectures and efficiency / Pruning and efficient tuning
Reborn Filters: Pruning Convolutional Neural Networks with Limited Data
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 |
|---|---|---|---|
| build compact filters from all original filters (reborn) instead of removing channels | channel pruning · CIFAR-10 VGGNet, N=500 | top-1 accuracy (%): 89.94±0.19 -> 93.03±0.19 | Helped |
| reborn filters vs DCP | ResNet-56 pruning · CIFAR-10, N=500 | top-1 accuracy (%): 83.19±0.47 -> 90.11±0.45 | Helped |
| inherit original filters vs training pruned network from scratch | pruned VGGNet-16 · ImageNet, N=1k | top-5 accuracy (%): 6.67±0.85 -> 89.40±0.03 | Helped |
| reborn filters vs DCP at N=100 | VGGNet-16 pruning · ImageNet, 2x FLOPs | top-5 accuracy (%): 78.74 -> 88.73 | Helped |
| use full training data instead of limited data | reborn filters vs DCP · ImageNet VGGNet-16 | top-5 accuracy (%): 91.68 -> 91.75 | No clear effect |
| add final fine-tuning after layer-wise reconstruction (ablation) | reborn filters · ImageNet, limited data | top-5 accuracy (%): about 0.1% improvement | No clear effect |
| increase acceleration rate to 3x | reborn filters vs Slimming · CIFAR-10 VGGNet, N=500 | accuracy (%): 80.15% -> 90.07% | Helped |
| increase sparsity weight lambda from 0.1 to 0.5 (ablation) | reborn filters · ImageNet VGGNet-16 conv1_2 | top-5 accuracy (%): 91.45 -> 85.45 | Hurt |
| use lambda below 0.3 (pruning 59%-78% channels) (ablation) | reborn filters · ImageNet VGGNet-16 conv1_2 | top-5 accuracy (%): 91.52 -> 91.40 | No clear effect |
Closest papers by meaning
- Rethinking the Pruning Criteria for Convolutional Neural Network (NeurIPS 2021 · similarity 0.91)
- Provable Filter Pruning for Efficient Neural Networks (ICLR 2020 · similarity 0.91)
- Pruning Filters for Efficient ConvNets (ICLR 2017 · similarity 0.91)
- UPDP: A Unified Progressive Depth Pruner for CNN and Vision Transformer (AAAI 2024 · similarity 0.89)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.89)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.88)
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.