Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning Filters for Efficient ConvNets
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 the smallest filters by L1 norm and retrain | VGG-16 · CIFAR-10 | Error (%): 6.75 -> 6.60 | Helped |
| Prune filters using the reported VGG-16 configuration | VGG-16 · CIFAR-10 | FLOP: $3.13 \times 10^8$ -> $2.06 \times 10^8$ | Helped |
| Train the pruned VGG-16 architecture from scratch rather than retrain inherited weights (ablation) | VGG-16-pruned-A · CIFAR-10 | Error (%): 6.60 -> 6.88 | Hurt |
| Skip sensitive layers while pruning filters | ResNet-56 · CIFAR-10 | Error (%): 6.96 -> 6.90 | Helped |
| Use stage-specific pruning rates and skip additional layers | ResNet-56 · CIFAR-10 | FLOP: $1.25 \times 10^8$ -> $9.09 \times 10^7$ | Helped |
| Train the pruned ResNet-56 architecture from scratch rather than retrain inherited weights (ablation) | ResNet-56-pruned-B · CIFAR-10 | Error (%): 6.94 -> 8.69 | Hurt |
| Prune filters and retrain while skipping sensitive layers | ResNet-110 · CIFAR-10 | Error (%): 6.47 -> 6.45 | Helped |
| Prune filters with the stage-specific configuration | ResNet-110 · CIFAR-10 | FLOP: $2.53 \times 10^8$ -> $1.55 \times 10^8$ | Helped |
| Prune first-layer residual filters rather than second-layer filters (ablation) | ResNet-34 · ILSVRC2012 | FLOP reduction: 24.2% -> 7.5% | Helped |
| Prune the third stage rather than the first residual-block layers (ablation) | ResNet-34 · ILSVRC2012 | Error (%): 27.83 -> 27.52 | Helped |
| Prune the smallest filters instead of random filters (ablation) | VGG-16 · CIFAR-10 | accuracy | Helped |
| Prune the largest filters instead of the smallest filters (ablation) | VGG-16 · CIFAR-10 | accuracy | Hurt |
Closest papers by meaning
- Interspace Pruning: Using Adaptive Filter Representations To Improve Training of Sparse CNNs (CVPR 2022 · similarity 0.92)
- Faster CNNs with Direct Sparse Convolutions and Guided Pruning (ICLR 2017 · similarity 0.92)
- Provable Filter Pruning for Efficient Neural Networks (ICLR 2020 · similarity 0.92)
- Gate Decorator: Global Filter Pruning Method for Accelerating Deep Convolutional Neural Networks (NeurIPS 2019 · similarity 0.92)
- Learning Filter Pruning Criteria for Deep Convolutional Neural Networks Acceleration (CVPR 2020 · similarity 0.92)
- Auto-Balanced Filter Pruning for Efficient Convolutional Neural Networks (AAAI 2018 · similarity 0.92)
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.