Research map / Architectures and efficiency / Pruning and efficient tuning
Winning the Lottery with Continuous Sparsification
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 Continuous Sparsification instead of magnitude-pruning ticket search | VGG-16 · CIFAR-10, sparsest matching subnetwork | test accuracy: 92.36% -> 93.35% | Helped |
| Use Continuous Sparsification instead of magnitude-pruning ticket search | ResNet-20 · CIFAR-10, sparsest matching subnetwork | weights remaining: 20.9% -> 12.3% | Helped |
| Use Continuous Sparsification instead of heuristic pruning | VGG-16 · CIFAR-10 pruning | sparsity within 2% test accuracy: 98.0% -> 99.6% | Helped |
| Use Continuous Sparsification instead of heuristic pruning | ResNet-20 · CIFAR-10 pruning | sparsity within 2% test accuracy: 86.0% -> 94.4% | Helped |
| Use Continuous Sparsification rather than IMP ticket search | ResNet-50 · ImageNet, retrained subnetwork at 90.0% sparsity | Top-1 accuracy: 73.6% -> 75.5% | Helped |
| Use a deterministic mask reparameterization rather than stochastic sparsification (ablation) | 6-layer CNN supermask search · CIFAR-10 | test accuracy at high sparsity: 25% or less remaining weights; 75% test accuracy | Helped |
| Replace stochastic sparsification with Continuous Sparsification (ablation) | Conv-6 ticket search · CIFAR-10 | ticket test accuracy and sparsity | Helped |
| Set the stochastic method's inference mask by sampling candidates and choosing the best (ablation) | Louizos-style stochastic ℓ0 pruning · VGG-16 and ResNet-20 pruning | test accuracy at high sparsity: above 90% | Helped |
| Apply both ℓ1 and ℓ2 regularization to scaling parameters instead of ℓ1 alone (ablation) | Network Slimming weight-wise scaling · VGG-16 and ResNet-20, CIFAR-10 | test performance | Hurt |
| Increase final inverse temperature to 200 from smaller tested values (ablation) | Continuous Sparsification · ResNet-20, CIFAR-10 ticket search | accuracy: 50 and 200 | Helped |
| Increase final inverse temperature to 200 (ablation) | Continuous Sparsification with s⁽⁰⁾=-0.2 or -0.3 · ResNet-20, CIFAR-10 ticket search | sparsity: 50 and 200 | Hurt |
| Vary initial mask value s⁽⁰⁾ (ablation) | Continuous Sparsification · ResNet-20, CIFAR-10 ticket search | sparsity: -0.3 and +0.3 | Helped |
Closest papers by meaning
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.94)
- Plant 'n' Seek: Can You Find the Winning Ticket? (ICLR 2022 · similarity 0.93)
- On the Existence of Universal Lottery Tickets (ICLR 2022 · similarity 0.92)
- Polynomially Over-Parameterized Convolutional Neural Networks Contain Structured Strong Winning Lottery Tickets (NeurIPS 2023 · similarity 0.92)
- Why Lottery Ticket Wins? A Theoretical Perspective of Sample Complexity on Sparse Neural Networks (NeurIPS 2021 · similarity 0.91)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.90)
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.