Research map / Learning theory and algorithms / Prediction and decision losses
WoodFisher: Efficient Second-Order Approximation for Neural Network Compression
Sidak Pal Singh, Dan Alistarh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 joint/global second-order pruning instead of global magnitude pruning | WoodFisher pruning · ImageNet, ResNet-50, 98% sparsity | Top-1 accuracy: 64.28 -> 65.55 | Improved |
| Use WoodFisher instead of STR | ResNet-50 gradual pruning · ImageNet, 80% sparsity | Top-1 accuracy: 76.19 -> 76.76 | Improved |
| Use WoodFisher instead of global magnitude pruning | MobileNetV1 gradual pruning · ImageNet, 89.00% sparsity | Top-1 accuracy: 63.02 -> 63.87 | Improved |
| Increase WoodFisher diagonal-block chunk size (ablation) | Joint WoodFisher · CIFAR10, ResNet-20 one-shot pruning | test accuracy | Improved |
| Use mini-batch-averaged gradients instead of individual-example gradients for Fisher (ablation) | WoodFisher · ImageNet, ResNet-50 one-shot pruning | Top-1 accuracy | Improved |
| Increase Fisher mini-batch size from 1 to 100 (ablation) | WoodFisher · ImageNet, ResNet-50 one-shot pruning at 30% sparsity | Top-1 accuracy: 55.81 ± 3.28 -> 75.73 ± 0.03 | Improved |
| Use 160 rather than 80 Fisher subsamples (ablation) | WoodFisher · ImageNet, ResNet-50 one-shot pruning at 30% sparsity, mini-batch size 100 | Top-1 accuracy: 75.73 ± 0.03 -> 75.77 ± 0.08 | No clear effect |
| Recompute the inverse-Hessian estimate during pruning without retraining (ablation) | WoodFisher · ImageNet, ResNet-50 one-shot pruning | Top-1 accuracy | Improved |
| Use model-sampled labels instead of empirical labels to compute Fisher gradients (ablation) | WoodFisher · ImageNet, ResNet-50 one-shot pruning | Top-1 accuracy: 76.10 ± 0.04 -> 76.16 ± 0.02; 75.22 ± 0.07 -> 75.31 ± 0.05; 69.21 ± 0.05 -> 69.29 ±0.20; 48.35 ± 0.22 -> 48.74 ±1.03; 5. | No clear effect |
| Include the first-order gradient term in pruning importance (WoodTaylor) (ablation) | WoodFisher pruning · MNIST, partially trained MLPNet | accuracy: about 5% | Improved |
| Include the first-order gradient term in pruning importance (WoodTaylor) (ablation) | WoodFisher pruning · CIFAR10, ResNet-20 | accuracy | Improved |
| Replace WoodFisher with diagonal-Fisher pruning (ablation) | One-shot pruning · CIFAR10, ResNet-20 above 30% sparsity | test accuracy | Hurt |
Closest papers by meaning
- AdaFisher: Adaptive Second Order Optimization via Fisher Information (ICLR 2025 · similarity 0.88)
- Error Feedback Can Accurately Compress Preconditioners (ICML 2024 · similarity 0.84)
- Exact, Tractable Gauss-Newton Optimization in Deep Reversible Architectures Reveal Poor Generalization (NeurIPS 2024 · similarity 0.84)
- Revisiting Scalable Hessian Diagonal Approximations for Applications in Reinforcement Learning (ICML 2024 · similarity 0.84)
- COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation (NeurIPS 2025 · similarity 0.84)
- Approximation to Smooth Functions by Low-Rank Swish Networks (ICML 2025 · similarity 0.84)
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.
