Research map / Architectures and efficiency / Pruning and efficient tuning
The Lottery Ticket Hypothesis: Finding Sparse, Trainable 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 |
|---|---|---|---|
| Find sparse subnetworks by iterative magnitude pruning and reset surviving weights to their original initialization | LeNet-300-100 · MNIST | test accuracy: improving by more than 0.3 percentage points when P_m = 13.5% | Helped |
| Find sparse subnetworks by iterative magnitude pruning and reset surviving weights to their original initialization | Conv-2 · CIFAR10 | test accuracy: 3.4 percentage points | Helped |
| Find sparse subnetworks by iterative magnitude pruning and reset surviving weights to their original initialization | Conv-4 · CIFAR10 | test accuracy: 3.5 | Helped |
| Find sparse subnetworks by iterative magnitude pruning and reset surviving weights to their original initialization | Conv-6 · CIFAR10 | test accuracy: 3.3 | Helped |
| Replace the original surviving weights with a fresh random initialization (ablation) | Winning-ticket masks · MNIST | test accuracy: half a percentage point | Hurt |
| Replace the original surviving weights with a fresh random initialization (ablation) | Winning-ticket masks · MNIST | early-stopping speed: 2.51x faster | Hurt |
| Use one-shot pruning instead of iterative pruning (ablation) | LeNet-300-100 · MNIST | test accuracy | Hurt |
| Add dropout rate 0.5 and iteratively prune (ablation) | Conv-2, Conv-4, and Conv-6 · CIFAR10 | test accuracy: up to an additional 2.3, 4.6, and 4.7 percentage points, respectively | Helped |
| Train VGG-19 at learning rate 0.1 without warmup (ablation) | VGG-19 · CIFAR10 | test accuracy | No clear effect |
| Add linear learning-rate warmup to VGG-19 at learning rate 0.1 (ablation) | VGG-19 · CIFAR10 | test accuracy: about one percentage point | Helped |
| Use learning rate 0.01 rather than 0.1 (ablation) | ResNet-18 · CIFAR10 | test accuracy: 89.5% ... falls short of ... 90.5% | Hurt |
| Add warmup and use learning rate 0.03 (ablation) | ResNet-18 · CIFAR10 | test accuracy: 90.5% | Helped |
Closest papers by meaning
- Why Lottery Ticket Wins? A Theoretical Perspective of Sample Complexity on Sparse Neural Networks (NeurIPS 2021 · similarity 0.94)
- Plant 'n' Seek: Can You Find the Winning Ticket? (ICLR 2022 · similarity 0.94)
- Winning the Lottery with Continuous Sparsification (NeurIPS 2020 · similarity 0.94)
- Winning Lottery Tickets in Deep Generative Models (AAAI 2021 · similarity 0.93)
- Find A Winning Sign: Sign Is All We Need to Win the Lottery (ICLR 2025 · similarity 0.92)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · similarity 0.92)
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.