Research map / Architectures and efficiency / Pruning and efficient tuning
Good Subnetworks Provably Exist: Pruning via Greedy Forward Selection
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 greedy forward selection pruning instead of the strongest listed baseline at the matched reported model size | ResNet34 on ImageNet2012 · ImageNet2012 | Top-1 accuracy: 72.5 -> 72.9 | Helped |
| Use greedy forward selection pruning instead of the listed strongest baseline at the matched reported FLOPs | MobileNetV2 on ImageNet2012 · ImageNet2012, 107M FLOPS | Top-1 accuracy: 65.0 -> 66.9 | Helped |
| Use greedy forward selection pruning instead of uniform pruning | MobileNetV3-Small on ImageNet2012 · ImageNet2012 | Top-1 accuracy: 65.4 -> 65.8 | Helped |
| Use greedy forward selection pruning instead of uniform pruning | ProxylessNet-Mobile on ImageNet2012 · ImageNet2012 | Top-1 accuracy: 72.9 -> 74.0 | Helped |
| Finetune pruned weights instead of retraining from random initialization (ablation) | MobileNetV2 · ImageNet, 220M FLOPs | Top-1 accuracy: 70.8 -> 71.6 | Helped |
| Finetune pruned weights instead of retraining from random initialization (ablation) | MobileNetV2 · ImageNet, 170M FLOPs | Top-1 accuracy: 69.0 -> 70.4 | Helped |
| Finetune pruned weights instead of retraining from random initialization (ablation) | MobileNetV3-Small · ImageNet, 49M FLOPs | Top-1 accuracy: 63.2 -> 65.8 | Helped |
| Use greedy forward selection rather than greedy backward elimination (ablation) | ResNet34 · ImageNet, 2.64G FLOPs (Forward) and 2.81G FLOPs (Backward) | Top-1 accuracy: 73.1 -> 73.5 | Helped |
| Use greedy forward selection rather than greedy backward elimination (ablation) | MobileNetV2 · ImageNet, 258M FLOPs (Forward) and 257M FLOPs (Backward) | Top-1 accuracy: 71.4 -> 71.9 | Helped |
| Prune from the full-size MobileNetV2 rather than MobileNetV2×0.5 (ablation) | MobileNetV2 · ImageNet, approximately 96M FLOPs pruned models | Top-1 accuracy: 64.9 -> 66.2 | Helped |
| Add greedy forward selection to a network previously pruned by network slimming | VGG19 · CIFAR-10 | Accuracy: 93.81 ± 0.14 -> 93.78 ± 0.16 | No clear effect |
| Add greedy forward selection to a network previously pruned by network slimming | PreResNet-164 · CIFAR-10 | Accuracy: 94.90 ± 0.04 -> 94.91 ± 0.06 | No clear effect |
Closest papers by meaning
- Greedy Optimization Provably Wins the Lottery: Logarithmic Number of Winning Tickets is Enough (NeurIPS 2020 · similarity 0.92)
- Pruning from Scratch (AAAI 2020 · similarity 0.91)
- Rethinking the Value of Network Pruning (ICLR 2019 · similarity 0.90)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.88)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.87)
- Device-Wise Federated Network Pruning (CVPR 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.