Research map / Architectures and efficiency / Pruning and efficient tuning
Channel Permutations for N:M Sparsity
Jeff Pool, Chong Yu
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 |
|---|---|---|---|
| channel permutations before 2:4 pruning | SqueezeNet v1.0 · ILSVRC12 | top-1 change vs dense: -4.01 -> 0.60 | Improved |
| channel permutations before 2:4 pruning | MobileNet v2 · ILSVRC12 | top-1 change vs dense: -1.98 -> 0.01 | Improved |
| channel permutations before 2:4 pruning | MobileNet v3 (Small) · ILSVRC12 | top-1 change vs dense: -2.73 -> 0.10 | Improved |
| channel permutations before 2:4 pruning | DenseNet-161 · ILSVRC12 | top-1 change vs dense: 0.82 -> 0.92 | Improved |
| channel permutations before 2:4 pruning | ResNet-50 · ILSVRC12 | top-1 change vs dense: 0.05 -> 0.13 | Improved |
| use bounded regressions escape phase (100 attempts) (ablation) | channel swap greedy search · random 64x128 matrices | efficacy (%): 46.5 -> 47.1 | Improved |
| optimize stripe groups of 2 instead of channel swap (ablation) | greedy phase · random 64x128 matrices | efficacy (%): 46.5 -> 47.7 | Improved |
| optimize stripe groups of 3 instead of 2 (ablation) | greedy phase · random 64x128 matrices | efficacy (%): 47.7 -> 51.9 | Improved |
| OSG(3) with BR(1000) vs channel swap on optimum finding (ablation) | search algorithm · 25 random 32x16 matrices | optimal found (count): 3 -> 25 | Improved |
| channel swap baseline vs 1000 random permutations (ablation) | search algorithm · random 64x128 matrices | efficacy (%): 8.0 -> 46.5 | Improved |
| search with channel swap only for MobileNet v2 (ablation) | Channel Swap permutations · ILSVRC12 | top-1 change vs dense: -1.26 | Mixed |
| OSG(3) BR(1000) instead of OSG(2) BR(100) in search time (ablation) | search algorithm · 2048x2048 matrices | search time: 29.99 -> - | Hurt |
Closest papers by meaning
- PermLLM: Learnable Channel Permutation for N:M Sparse Large Language Models (NeurIPS 2025 · similarity 0.91)
- Learnable Permutation for Structured Sparsity on Transformer Models (AAAI 2026 · similarity 0.88)
- Learning Best Combination for Efficient N:M Sparsity (NeurIPS 2022 · similarity 0.88)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.87)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.85)
- TSENOR: Highly-Efficient Algorithm for Finding Transposable N:M Sparse Masks (NeurIPS 2025 · similarity 0.85)
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.
