Research map / Architectures and efficiency / Pruning and efficient tuning
Greedy Optimization Provably Wins the Lottery: Logarithmic Number of Winning Tickets is Enough
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 |
|---|---|---|---|
| Combine local and global imitation instead of global only (ablation) | Greedy layer-wise pruning · ImageNet ResNet34 2.2G FLOPs | top-1 accuracy: 73.2 -> 73.5 | Helped |
| Combine local and global imitation instead of global only (ablation) | Greedy layer-wise pruning · ImageNet MobileNetV2 245M FLOPs | top-1 accuracy: 72.1 -> 72.2 | Helped |
| Prune ResNet34 with the proposed method instead of GFS (~2.2G FLOPs) | Ours · ImageNet ResNet34 | top-1 accuracy / FLOPs: 73.5 at 2.64G -> 73.5 at 2.20G | Helped |
| Prune ResNet34 harder (1.90G FLOPs) versus GFS | Ours · ImageNet ResNet34 | top-1 accuracy: 72.9 -> 73.3 | Helped |
| Prune MobileNetV2 with Ours instead of GFS (~220M FLOPs) | Ours · ImageNet MobileNetV2 | top-1 accuracy: 71.6 -> 71.7 | Helped |
| Prune MobileNetV2 to 170M FLOPs: Ours vs GFS | Ours · ImageNet MobileNetV2 | top-1 accuracy: 70.4 -> 70.5 | No clear effect |
| Prune MobileNetV3-Small: Ours vs GFS | Ours · ImageNet MobileNetV3-Small | top-1 accuracy: 65.8 -> 66.4 | Helped |
| Prune MobileNetV3-Small: Ours vs Uniform | Ours · ImageNet MobileNetV3-Small | top-1 accuracy: 65.4 -> 66.4 | Helped |
| Prune DGCNN to 0.5x width with Ours instead of uniform width scaling | Ours · ModelNet40 | accuracy: 92.3 -> 92.7 | Helped |
| Prune DGCNN to 0.25x width with Ours | Ours · ModelNet40 | accuracy: 91.8 -> 92.5 | Helped |
| Ours vs S3D at similar size | Ours · ModelNet40 | forward time (ms): 42.06 -> 37.43 | Helped |
| Ours vs DGCNN 0.5x forward time | Ours · ModelNet40 | forward time (ms): 38.90 -> 28.06 | Helped |
Closest papers by meaning
- Good Subnetworks Provably Exist: Pruning via Greedy Forward Selection (ICML 2020 · similarity 0.92)
- Winning the Lottery with Continuous Sparsification (NeurIPS 2020 · similarity 0.90)
- Rethinking the Value of Network Pruning (ICLR 2019 · similarity 0.90)
- Picking Winning Tickets Before Training by Preserving Gradient Flow (ICLR 2020 · similarity 0.90)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.90)
- No Free Prune: Information-Theoretic Barriers to Pruning at Initialization (ICML 2024 · similarity 0.88)
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.