Research map / Architectures and efficiency / Pruning and efficient tuning
Accelerate CNN via Recursive Bayesian Pruning
Yuefu Zhou, Ya Zhang, Yanfeng Wang, Qi Tian
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 |
|---|---|---|---|
| prune layer by layer with inter-layer dependency (RBP) instead of all layers independently (IBP) (ablation) | RBP · CIFAR10 VGG16 | compression rate and FLOPs reduction: 13.3x -> 39.1x CR, 2.3x -> 3.1x ... 3.5x FLOPs; error 8.3 -> 9.0 | Improved |
| apply BC layer by layer (RBC) instead of simultaneously (ablation) | BC · CIFAR10 VGG16 | compression rate, FLOPs, error: 18.6x -> 25.7x CR, 2.6x -> 3.1x FLOPs, error 9.0 -> 9.5 | Mixed |
| use RBP instead of BC/SBP on VGG16 | VGG16 CIFAR10 · CIFAR10 | FLOPs reduction: 3.2x/2.6x -> 3.5x at error 9.0 | Improved |
| use RBP instead of DDS on VGG16 | VGG16 ILSVRC · ILSVRC2012 | FLOPs reduction and top-1 error: 4.0x -> 5.0x FLOPs, top-1 error 31.5 -> 31.6 | Mixed |
| use RBP instead of ThiNet on VGG16 | VGG16 ILSVRC · ILSVRC2012 | FLOPs reduction and top-1 error: 3.2x -> 5.0x FLOPs, top-1 error 30.2 -> 31.6 | Mixed |
| use RBP instead of CP on VGG16 | VGG16 ILSVRC · ILSVRC2012 | FLOPs and top-1 error: 4.4x -> 5.0x FLOPs, top-1 error 32.2 -> 31.6 | Improved |
| use ResNet-adaptive RRBP (skip down-sampling blocks, parallel pruning) instead of plain RBP | ResNet50 · ILSVRC2012 | top-1 error: 28.9 -> 27.0 top-1 error, 10.0 -> 9.0 top-5 | Improved |
| use RRBP instead of ThiNet-50 | ResNet50 · ILSVRC2012 | top-1 and top-5 error: 29.0 -> 27.0, 10.0 -> 9.0 at similar FLOPs (2.3x vs 2.2x) | Improved |
| use RBP/RRBP instead of DDS(26) | ResNet50 · ILSVRC2012 | FLOPs and top-1 error: 1.7x -> 2.3x/2.2x FLOPs; top-1 error 28.2 -> 28.9/27.0 | Improved |
| prune with RBP on GPU | VGG16 ILSVRC · ILSVRC2012 | inference speed-up: 2.6x time, 3.6x storage, top-5 error +2.2 | Improved |
| prune ResNet50 with RRBP vs RBP on GPU | ResNet50 · ILSVRC2012 | top-5 error increase: +2.9 -> +1.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- ThiNet — ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression (ICCV 2017) · better on 2 of 4 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
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.91)
- Bayesian Optimization with Clustering and Rollback for CNN Auto Pruning (ECCV 2022 · similarity 0.89)
- BMRS: Bayesian Model Reduction for Structured Pruning (NeurIPS 2024 · similarity 0.89)
- Accelerating CNN Training by Pruning Activation Gradients (ECCV 2020 · similarity 0.89)
- Faster CNNs with Direct Sparse Convolutions and Guided Pruning (ICLR 2017 · similarity 0.89)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · 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.
