Research map / Architectures and efficiency / Pruning and efficient tuning
Gradient Flow in Sparse Neural Networks and How Lottery Tickets Win
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 sparsity-aware initialization instead of masked dense initialization | Sparse VGG16 trained from scratch · ImageNet-2012, VGG16 (80% sparse) | Top-1 accuracy: 51.81 -> 62.52 | Helped |
| Use sparsity-aware initialization instead of masked dense initialization | Sparse LeNet5 trained from scratch · MNIST, LeNet5 (98% sparse) | test accuracy: 56.51$\pm$36.90 -> 90.00$\pm$4.04 | Helped |
| Use sparsity-aware initialization instead of masked dense initialization | Sparse ResNet50 · ImageNet-2012, ResNet50 (80% sparse) | Top-1 accuracy: 70.58$\pm$0.18 -> 70.63$\pm$0.22 | No clear effect |
| Use Liu initialization instead of masked dense initialization | Sparse LeNet5 trained from scratch · MNIST, LeNet5 (95% sparse) | test accuracy: 62.99$\pm$42.16 -> 96.64$\pm$0.83 | Helped |
| Use sparsity-aware initialization instead of masked dense initialization | Sparse LeNet5 trained from scratch · MNIST, LeNet5 (95% sparse) | test accuracy: 62.99$\pm$42.16 -> 97.70$\pm$0.09 | Helped |
| Use RigL connectivity updates rather than fixed sparse training | Sparse models with masked dense initialization · MNIST, LeNet5 (95% sparse) | test accuracy: 62.99$\pm$42.16 -> 80.82$\pm$34.74 | Helped |
| Choose new RigL connections by smallest rather than largest gradient (ablation) | RigL · ImageNet-2012, ResNet50 | test accuracy: 73.83$\pm$0.12 | Hurt |
| Choose new RigL connections by smallest rather than largest gradient (ablation) | RigL · MNIST, LeNet5 | test accuracy: 92.71$\pm$7.67 | Hurt |
| Use SET random connection growth instead of RigL gradient-based growth (ablation) | Dynamic sparse training · Sparse neural networks | gradient flow | Hurt |
| Use RigL gradient-based growth instead of SET random growth | Dynamic sparse training · Sparse neural networks | gradient flow | Helped |
| Train from a lottery-ticket initialization rather than scratch | Sparse LeNet5 · MNIST, LeNet5 | fractional disagreement with pruning solution: 0.0278 -> 0.0089 | Helped |
| Train an ensemble from lottery-ticket rather than random sparse initializations | Sparse models · MNIST and ImageNet-2012 | ensemble accuracy gain: +0.06% and +0.54% vs +0.96% and +2.89% | Mixed |
Closest papers by meaning
- Lottery Tickets in Evolutionary Optimization: On Sparse Backpropagation-Free Trainability (ICML 2023 · similarity 0.91)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.89)
- Find A Winning Sign: Sign Is All We Need to Win the Lottery (ICLR 2025 · similarity 0.88)
- Sparse Training from Random Initialization: Aligning Lottery Ticket Masks using Weight Symmetry (ICML 2025 · similarity 0.88)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · similarity 0.88)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (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.