Research map / Architectures and efficiency / RNNs and normalization
Dynamic Sparse Graph for Efficient Deep Learning
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 dimension-reduction search rather than random neuron selection (ablation) | DSG · CIFAR10 | accuracy | Helped |
| Estimate activations with dimension-reduction search instead of oracle top-k selection (ablation) | DSG · CIFAR10 | accuracy | No clear effect |
| Use double masks, including a second mask after BN (ablation) | DSG · CIFAR10 | accuracy | Helped |
| Remove batch normalization (ablation) | DSG · CIFAR10 | accuracy | Hurt |
| Increase sparsity beyond 80% (ablation) | DSG · small- and medium-scale model benchmarks | accuracy | Hurt |
| Use lower epsilon for dimension-reduction search (ablation) | DSG · CIFAR10 | accuracy | Helped |
| Use wider, shallower WRN rather than deeper, slimmer ResNet (ablation) | DSG · CIFAR10 validation set | accuracy: 1% | Mixed |
| Replace DSG with a smaller dense model having equivalent MACs (ablation) | VGG8 and other evaluated networks · CIFAR10 and ImageNet | accuracy | Hurt |
| Use DSG rather than the dense baseline | VGG8 · CIFAR10 | training time | Helped |
| Use DSG with dynamic vector-wise sparsity rather than dense GEMM | VGG8 CPU execution · VGG8 layers | speedup: 0.6x, 1.6x, and 2.7x | Mixed |
| Use DSG at sparsity below 60% (ablation) | DSG · small- and medium-scale model benchmarks | accuracy | No clear effect |
Closest papers by meaning
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.88)
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities (ICML 2024 · similarity 0.88)
- Dynamic Sparse Training of Diagonally Sparse Networks (ICML 2025 · similarity 0.88)
- Dynamic Sparse Training with Structured Sparsity (ICLR 2024 · similarity 0.88)
- Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · similarity 0.87)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.87)
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.