Research map / Architectures and efficiency / Pruning and efficient tuning
Neural Network Pruning Denoises the Features and Makes Local Connectivity Emerge in Visual Tasks
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 |
|---|---|---|---|
| iterative magnitude pruning (IMP) instead of no pruning | FCN 3x1024 on ImageNet32 · ImageNet32 | highest validation accuracy: from less than 9% to more than 12% | Helped |
| reduce training set size to 2-5% of original (ablation) | IMP on FCN · ImageNet32 subsets | gain from IMP over dense accuracy: improvement decreases until pruned network only maintains dense accuracy | Hurt |
| use semantic 10-class task instead of random 10-class grouping (ablation) | IMP on FCN · ImageNet32 10-macro-class tasks | validation accuracy and pruning benefit: random at best 13%, semantic surpasses 36% | Helped |
| one-shot pruning to high sparsity instead of iterative pruning (ablation) | magnitude pruning on FCN · ImageNet32 | mask locality and structure | Hurt |
| SNIP pruning at initialization instead of IMP (ablation) | FCN on ImageNet32 · ImageNet32 | mask resemblance to final weights and locality | Hurt |
| SynFlow data-free pruning instead of IMP (ablation) | FCN on ImageNet32 · ImageNet32 | mask locality and structure | Hurt |
Closest papers by meaning
- On the Predictability of Pruning Across Scales (ICML 2021 · similarity 0.91)
- Pruning from Scratch (AAAI 2020 · similarity 0.90)
- Pruning neural networks without any data by iteratively conserving synaptic flow (NeurIPS 2020 · similarity 0.90)
- When To Prune? A Policy Towards Early Structural Pruning (CVPR 2022 · similarity 0.90)
- Rethinking the Value of Network Pruning (ICLR 2019 · similarity 0.90)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 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.