Research map / Architectures and efficiency / Pruning and efficient tuning
A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations
Hongrong Cheng, Miao Zhang, Qinfeng Shi
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 |
|---|---|---|---|
| use unstructured instead of structured pruning (SNIP) (ablation) | SNIP on VGG-16 · CIFAR-100 (50% pruned) | Top-1 accuracy: 70.68 -> 72.67 | Improved |
| use unstructured instead of structured pruning (GraSP) (ablation) | GraSP on VGG-16 · CIFAR-10 (40% pruned) | Top-1 accuracy: 93.04 -> 93.18 | Improved |
| use unstructured instead of structured pruning at 10% on GraSP (ablation) | GraSP on VGG-16 · CIFAR-10 (10% pruned) | Top-1 accuracy: 93.67 -> 93.58 | No clear effect |
| use unstructured instead of 2:4 semi-structured pruning (ablation) | SparseGPT on OPT-125M · WikiText2 | perplexity: 45.51 -> 33.17 | Improved |
| use 2:4 semi-structured instead of structured pruning (ablation) | K-OBD on OPT-125M · WikiText2 | perplexity: 75.95 -> 68.74 | Improved |
| use unstructured instead of structured pruning (ablation) | LLM-Surgeon on OPT-6.7B · WikiText2 | perplexity: 14.90 -> 10.97 | Improved |
| use iterative instead of one-shot pruning (ablation) | SynFlow, Magnitude, SparseGPT · CIFAR-10, PTB | accuracy / perplexity | Improved |
| increase the number of pruning iterations (ablation) | pruning of OPT-1.3B · WikiText2 | perplexity | Improved |
| use data-driven SNIP instead of data-free Random pruning (ablation) | VGG-16 · CIFAR-10 (90% pruned) | Top-1 accuracy: 90.68 -> 93.40 | Improved |
| use data-free SynFlow instead of data-driven GraSP (ablation) | VGG-16 · CIFAR-10 (90% pruned) | Top-1 accuracy: 92.51 -> 93.26 | Improved |
| use data-driven LoRAPruner instead of data-free L2-norm (ablation) | LLaMA-7B (20% pruned) · zero-shot common sense avg | accuracy: 66.84 -> 69.42 | Improved |
| use LoRAPruner instead of Wanda at 50% ratio (ablation) | LLaMA-7B · zero-shot common sense avg | accuracy: 50.60 -> 59.07 | Improved |
Closest papers by meaning
- Computation and Memory-Efficient Model Compression with Gradient Reweighting (NeurIPS 2025 · similarity 0.94)
- Structured Pruning for Deep Convolutional Neural Networks: A Survey (TPAMI 2023 · similarity 0.94)
- Learning Pruning-Friendly Networks via Frank-Wolfe: One-Shot, Any-Sparsity, And No Retraining (ICLR 2022 · similarity 0.92)
- What Makes a Good Prune? Maximal Unstructured Pruning for Maximal Cosine Similarity (ICLR 2024 · similarity 0.92)
- Convolutional Neural Network Pruning With Structural Redundancy Reduction (CVPR 2021 · similarity 0.92)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.91)
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.
