Research map / Architectures and efficiency / Pruning and efficient tuning
SInGE: Sparsity via Integrated Gradients Estimation of Neuron Relevance
Edouard YVINEC, Arnaud Dapogny, Matthieu Cord, Kevin Bailly
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 L2 norm instead of L1 for magnitude criterion (ablation) | magnitude pruning · CIFAR10 ResNet-56, 75% params removed | top-1 accuracy: 42.01 -> 42.35 | No clear effect |
| use gradient criterion instead of magnitude (ablation) | pruning criterion · CIFAR10 ResNet-56, 75% params removed | top-1 accuracy: 42.35 -> 77.68 | Improved |
| combine magnitude x gradient instead of gradient only (ablation) | pruning criterion · CIFAR10 ResNet-56, 85% params removed | top-1 accuracy: 28.31 -> 90.28 | Improved |
| integrate gradient along path to zero (IG) instead of magnitude x gradient (ablation) | pruning criterion · CIFAR10 ResNet-56, 85% params removed | top-1 accuracy: 90.28 -> 92.80 | Improved |
| integrate gradient along path to zero (IG) instead of magnitude x gradient (ablation) | pruning criterion · CIFAR10 ResNet-56, 90% params removed | top-1 accuracy: 10.00 -> 85.38 | Improved |
| keep magnitude term in integrated criterion (IG vs SG) (ablation) | pruning criterion · CIFAR10 ResNet-56, 90% params removed | top-1 accuracy: 77.20 -> 85.38 | Improved |
| entwine pruning and fine-tuning instead of post-pruning fine-tuning (ablation) | SInGE · CIFAR10 ResNet-56, 90% params removed, 1000 steps | top-1 accuracy: 77.2 -> 85.38 | Improved |
| entwine pruning and fine-tuning instead of post-pruning (ablation) | SInGE · CIFAR10 ResNet-56, 85% params removed, 5000 steps | top-1 accuracy: 93.13 -> 93.31 | Improved |
| increase fine-tuning steps from 1000 to 5000 (entwined) (ablation) | SInGE · CIFAR10 ResNet-56, 90% params removed | top-1 accuracy: 85.38 -> 90.02 | Improved |
| increase fine-tuning steps from 1000 to 5000 (entwined) (ablation) | SInGE · CIFAR10 ResNet-56, 85% params removed | top-1 accuracy: 93.18 -> 93.31 | Improved |
| SInGE vs HAP at lossless accuracy | SInGE · CIFAR10 ResNet-56 (93.5 acc) | % parameters removed: 76.2 -> 84.3 | Improved |
| SInGE vs MDP at 40% goal | SInGE · ImageNet MobileNetV2 | top-1 accuracy: 69.58 -> 70.47 | Improved |
Closest papers by meaning
- Importance Estimation for Neural Network Pruning (CVPR 2019 · similarity 0.90)
- NISP: Pruning Networks Using Neuron Importance Score Propagation (CVPR 2018 · similarity 0.89)
- Isomorphic Pruning for Vision Models (ECCV 2024 · similarity 0.89)
- Prospect Pruning: Finding Trainable Weights at Initialization using Meta-Gradients (ICLR 2022 · similarity 0.88)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.87)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 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.
