Research map / Architectures and efficiency / CNN design
Learning Structured Sparsity in Deep Neural Networks
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 |
|---|---|---|---|
| Apply filter+channel SSL to LeNet (LeNet 2) | LeNet · MNIST | error: 0.9% -> 0.8% | No clear effect |
| Stronger filter-wise sparsity (LeNet 3) (ablation) | LeNet · MNIST | error: 0.8% -> 1.0% | Hurt |
| Shape-wise SSL, size 25 to 7 (LeNet 5) | LeNet · MNIST | conv1 speedup: 1.00x -> 5.19x | Helped |
| Row/column sparsity SSL (ConvNet 2) | ConvNet · CIFAR-10 | error: 17.9% -> 17.9% | No clear effect |
| Row/column sparsity SSL (ConvNet 3) | ConvNet · CIFAR-10 | error: 17.9% -> 16.9% | Helped |
| Depth-wise SSL to 18 layers (SSL-ResNet-18) | ResNet-20 · CIFAR-10 | error: 8.82% -> 7.40% | Helped |
| Depth-wise SSL to 14 layers | ResNet-20 · CIFAR-10 | error | Helped |
| Preserve fewer layers via stronger depth regularization (ablation) | SSL-ResNet · CIFAR-10 | error | Hurt |
| Raise depth from 18 to 20 in SSL-ResNet (ablation) | SSL-ResNet · CIFAR-10 | error | Hurt |
| 2D-filter-wise sparsity with parameters retained | AlexNet · ImageNet | top-1 error: 42.63% -> 41.69% | Helped |
| Shape-wise sparsity with parameters retained | AlexNet · ImageNet | top-1 error: 42.63% -> 41.83% | Helped |
| SSL instead of l1-norm for GPU speedup | AlexNet · ImageNet | average layer-wise speedup (GPU): 0.9x -> 3.1x | Helped |
Closest papers by meaning
- SCHTs: A Semi-Structured Dynamic Sparse Training Framework for Hardware-Efficient Deep Learning (NeurIPS 2026 · similarity 0.88)
- Structured Sparsity Learning for Efficient Video Super-Resolution (CVPR 2023 · similarity 0.87)
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.87)
- Learning N:M Fine-grained Structured Sparse Neural Networks From Scratch (ICLR 2021 · similarity 0.87)
- Structured Gradient-based Interpretations via Norm-Regularized Adversarial Training (CVPR 2024 · similarity 0.85)
- Differentiable Sparsity via $D$-Gating: Simple and Versatile Structured Penalization (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.