Research map / Architectures and efficiency / Pruning and efficient tuning
Fast as CHITA: Neural Network Pruning with Combinatorial Optimization
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 |
|---|---|---|---|
| multi-stage procedure (CHITA++) instead of single-stage CHITA (ablation) | MLPNet pruning · MNIST, sparsity 0.98 | test accuracy: 46.25 to 90.73 | Helped |
| multi-stage procedure (CHITA++) instead of single-stage CHITA (ablation) | ResNet20 pruning · CIFAR10, sparsity 0.9 | test accuracy: 15.60 to 79.32 | Helped |
| single-stage CHITA vs. CBS/WF/MP one-shot | MLPNet, ResNet20, MobileNetV1 one-shot pruning · MNIST/CIFAR10/ImageNet one-shot | test accuracy | No clear effect |
| multi-stage CHITA++ vs. existing one-shot methods | MLPNet, ResNet20, MobileNetV1 · one-shot pruning at high sparsity | test accuracy: e.g., MLPNet 0.98: 90.73 vs 55.45 (CBS) | Helped |
| plug CHITA into gradual pruning (block-wise BSO) | MobileNetV1 gradual pruning · ImageNet, 89% sparsity | pruned accuracy: 67.68 vs 63.87 (WoodFisher) | Helped |
| plug CHITA into gradual pruning | ResNet50 gradual pruning · ImageNet, 98% sparsity | pruned accuracy: 69.80 vs 65.55 (WoodFisher) | Helped |
| ridge regularization term (lambda) added to the objective (ablation) | MLPNet pruning · MNIST test accuracy | test accuracy: approximately 3% increase | Helped |
| block-wise approximation (block size 1e4) instead of full-matrix (ablation) | MobileNetV1 one-shot pruning · ImageNet, sparsity 0.8 | test accuracy: from 0.2% to near 30% | Helped |
| scaled first-order term (alpha=1/m) instead of unscaled (alpha=1) (ablation) | MLPNet with CHITA++ at 0.95 sparsity · MNIST, fisher batch size m>1 | test accuracy | Helped |
| exponential vs linear sparsity schedule in CHITA++ (ablation) | ResNet20 CHITA++ with 15 stages · CIFAR10, 90% sparsity | test accuracy | Mixed |
| adding the first-order gradient term in single-stage problem | CHITA single-stage · MNIST/CIFAR10 | test accuracy | No clear effect |
Closest papers by meaning
- “Learning-Compression” Algorithms for Neural Net Pruning (CVPR 2018 · similarity 0.91)
- Sparsest Models Elude Pruning: An Exposé of Pruning’s Current Capabilities (ICML 2024 · similarity 0.89)
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.88)
- How Sparse Can We Prune A Deep Network: A Fundamental Limit Perspective (NeurIPS 2024 · similarity 0.88)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.88)
- EPSD: Early Pruning with Self-Distillation for Efficient Model Compression (AAAI 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.