Research map / Architectures and efficiency / Pruning and efficient tuning
GDP: Stabilized Neural Network Pruning via Gates With Differentiable Polarization
Yi Guo, Huan Yuan, Jianchao Tan, Zhangyang Wang, Sen Yang, Ji Liu
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 |
|---|---|---|---|
| introduce separate learnable gate parameters alpha instead of using conv kernel norms as gate input (ablation) | GDP · ImageNet MobileNet-V2 | top-1 accuracy | Improved |
| use gates g(alpha) in forward pass instead of naive smooth-L0 regularization on weights (ablation) | GDP · ImageNet MobileNet-V2 | top-1 accuracy | Improved |
| prune from a well-trained model instead of from scratch (ablation) | GDP · ImageNet MobileNet-V2 | top-1 accuracy | Improved |
| use proximal l1 relaxation instead of proximal l0 (ablation) | GDP optimizer · pruning | stability | Improved |
| vary decay rate of epsilon from 0.86 to 0.96 (ablation) | GDP · ImageNet MobileNet-V2 | top-1 accuracy / compression | No clear effect |
| vary initial epsilon value (ablation) | GDP · ImageNet MobileNet-V2 | top-1 accuracy / compression | No clear effect |
| prune DeepLabv3+ ResNet-50 to 38.4% FLOPs | GDP · Pascal VOC 2012 val | mIoU: 0.774 -> 0.779 | Improved |
| prune DeepLabv3+ ResNet-50 to 5.2% FLOPs | GDP · Pascal VOC 2012 val | mIoU: 0.774 -> 0.666 | Hurt |
| prune VGG-16 with GDP vs unpruned baseline | GDP · CIFAR-10 VGG-16 | accuracy: 93.89 -> 93.99 | Improved |
| prune ResNet-56 to 34.36% FLOPs | GDP · CIFAR-10 ResNet-56 | accuracy: 93.9 -> 93.55 | Hurt |
| prune ResNet-56 to 46.65% FLOPs | GDP · CIFAR-10 ResNet-56 | accuracy: 93.9 -> 93.97 | Improved |
| prune MobileNet-V2 to 53.78% FLOPs | GDP · CIFAR-10 MobileNet-V2 | accuracy: 94.89 -> 95.15 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- HRank — HRank: Filter Pruning Using High-Rank Feature Map (CVPR 2020) · better on 2 of 2 shared results
- SCOP — SCOP: Scientific Control for Reliable Neural Network Pruning (NeurIPS 2020) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Disentangled Differentiable Network Pruning (ECCV 2022 · similarity 0.90)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.90)
- Revisiting Random Channel Pruning for Neural Network Compression (CVPR 2022 · similarity 0.90)
- Discrimination-aware Channel Pruning for Deep Neural Networks (NeurIPS 2018 · similarity 0.89)
- Operation-Aware Soft Channel Pruning using Differentiable Masks (ICML 2020 · similarity 0.89)
- Learning Pruning-Friendly Networks via Frank-Wolfe: One-Shot, Any-Sparsity, And No Retraining (ICLR 2022 · similarity 0.89)
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.
