Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning Convolutional Neural Networks for Resource Efficient Inference
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 |
|---|---|---|---|
| Use Oracle-abs rather than Oracle-loss for pruning (ablation) | Greedy feature-map pruning on Birds-200 · Birds-200 | training accuracy | Helped |
| Add layer-wise ℓ2-normalization to criterion scores (ablation) | Taylor criterion · VGG-16 / Birds-200, all layers | Spearman rank correlation: 0.14 -> 0.73 | Helped |
| Add layer-wise ℓ2-normalization to criterion scores (ablation) | Taylor criterion · AlexNet / Flowers-102, all layers | Spearman rank correlation: 0.37 -> 0.75 | Helped |
| Increase updates between pruning iterations from 100 to 1000 (ablation) | Taylor and random pruning on AlexNet / ImageNet · ImageNet | accuracy | Helped |
| Add five epochs of fine-tuning after pruning to 11.5 GFLOPs (ablation) | VGG-16 · ImageNet top-5 validation | accuracy: 83% -> 87% | Helped |
| Add five epochs of fine-tuning after pruning to 8.0 GFLOPs (ablation) | VGG-16 · ImageNet top-5 validation | accuracy: 77.8% -> 84.5% | Helped |
| Add FLOPs regularization to Taylor pruning (ablation) | VGG-16 pruning · Birds-200 | accuracy at a given number of operations | Helped |
| Replace iterative pruning with high ℓ2 regularization and threshold-based pruning (ablation) | VGG-16 last convolutional layer · Birds-200, pruning 85% or more of feature maps | test accuracy | Hurt |
| Linearly combine Taylor and mean-activation criteria (ablation) | Pruning saliency estimation · Birds-200 and Flowers-102 | Spearman rank correlation | No clear effect |
| Increase fine-tuning updates between pruning iterations (ablation) | Taylor pruning on AlexNet / Flowers-102 · Flowers-102 | accuracy | Helped |
| Prune with Taylor criterion instead of random or minimum-weight ranking | AlexNet · ImageNet, 100 updates between pruning iterations | accuracy: 0%-4% | Helped |
Closest papers by meaning
- RePr: Improved Training of Convolutional Filters (CVPR 2019 · similarity 0.90)
- Neural Network Pruning With Residual-Connections and Limited-Data (CVPR 2020 · similarity 0.90)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.89)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.88)
- OSSCAR: One-Shot Structured Pruning in Vision and Language Models with Combinatorial Optimization (ICML 2024 · similarity 0.88)
- MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (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.