Research map / Optimization / Implicit bias and generalization
Lottery Tickets in Evolutionary Optimization: On Sparse Backpropagation-Free Trainability
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 an IMP-derived initialization instead of random pruning | Evolution strategies · 12 control and vision tasks | task performance | Helped |
| Use SNR-based pruning instead of magnitude-based IMP | Evolution strategies · Control and vision tasks | task performance | Mixed |
| Use SNR-pruned ES initializations instead of sparsity-matched GD-IMP initializations | Sparse network training · Control and vision tasks at high sparsity | task performance | Helped |
| Transfer an SNR-pruned initialization from a related task instead of random pruning | Evolution strategies · Ant Mass 10 to Ant Mass 15; F-MNIST to MNIST | task performance | Helped |
| Transfer an SNR-pruned initialization between evolution strategies instead of random pruning | Evolution strategies · Ant and F-MNIST transfer settings | task performance | Helped |
| Transfer an SNR-pruned ES initialization to gradient-based training instead of random pruning | GD-based training · Ant and F-MNIST transfer settings | task performance | Mixed |
| Increase sparsity during IMP | GD-IMP · Vision tasks | test accuracy | Hurt |
| Increase sparsity during IMP | ES-based training · Vision tasks | local-optimum sharpness | No clear effect |
| Use the final evolved weight magnitude as the pruning heuristic instead of other tested heuristics (ablation) | ES iterative pruning · Four tasks across three ES | task performance | No clear effect |
| Vary the network initialization scheme (ablation) | ES sparse trainability · Four tasks across three ES | task performance | No clear effect |
| Use movement, magnitude increase, or smallest-initial-value pruning instead of final-magnitude pruning (ablation) | ES iterative pruning · Four tasks across three ES | task performance | No clear effect |
| Account for search-distribution standard deviation when pruning, using SNR rather than weight magnitude alone (ablation) | ES iterative pruning · Fetch and F-MNIST | task performance | Mixed |
Closest papers by meaning
- Gradient Flow in Sparse Neural Networks and How Lottery Tickets Win (AAAI 2022 · similarity 0.91)
- The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks (ICLR 2019 · similarity 0.90)
- Validating the Lottery Ticket Hypothesis with Inertial Manifold Theory (NeurIPS 2021 · similarity 0.89)
- Find A Winning Sign: Sign Is All We Need to Win the Lottery (ICLR 2025 · similarity 0.88)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.87)
- Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win (ICML 2025 · 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.