Research map / Architectures and efficiency / Quantization
Generalizable Mixed-Precision Quantization via Attribution Rank Preservation
Ziwei Wang, Han Xiao, Jiwen Lu, Jie Zhou
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 |
|---|---|---|---|
| search policy with attribution rank preservation on CIFAR-10 instead of HAQ/EdMIPS search on CIFAR-10 | mixed-precision quantization · ImageNet MobileNet-V2 (~7-8 GBOPs) | top-1 accuracy: 62.7 -> 70.4 | Improved |
| add GMPQ rank-preservation to policy found on CIFAR-10 | EdMIPS-C · ImageNet ResNet50 (~16 GBOPs) | top-1 accuracy: 65.6 -> 73.6 | Improved |
| add GMPQ rank-preservation to policy found on CIFAR-10 | EdMIPS-C · ImageNet ResNet18 (~7 GBOPs) | top-1 accuracy: 59.1 -> 67.8 | Improved |
| search on CIFAR-10 with GMPQ instead of on ImageNet with EdMIPS | ResNet50 · ImageNet (~16 GBOPs) | search cost (GPU hours): 26.5 -> 3.4 | Improved |
| search on CIFAR-10 with GMPQ instead of on ImageNet with HAQ | MobileNet-V2 · ImageNet (~7-8 GBOPs) | search cost (GPU hours): 51.1 -> 2.6 | Improved |
| use GMPQ instead of HAWQ | ResNet18 · ImageNet (~28-34 GBOPs) | top-1 accuracy: 68.5 -> 70.2 | Improved |
| use GMPQ with policy searched on CIFAR-10 vs HAQ-C | SSD VGG16 · PASCAL VOC07 (~420 GBOPs) | mAP: 66.4 -> 69.2 | Improved |
| use GMPQ with policy searched on CIFAR-10 vs HAQ-C | Faster R-CNN ResNet18 · PASCAL VOC07 (~290 GBOPs) | mAP: 70.4 -> 73.4 | Improved |
| use GMPQ instead of HAQ searched on VOC | SSD VGG16 · PASCAL VOC07 (~800 GBOPs) | mAP: 70.9 -> 70.5 | Hurt |
| use capacity-aware p instead of fixed p in attribution imitation (ablation) | GMPQ ResNet18 · ImageNet accuracy-complexity trade-off | accuracy-complexity curve | Improved |
| use medium fixed p (versus small or large p) (ablation) | GMPQ · ImageNet ResNet18 | accuracy-complexity curve | Improved |
| use medium generalization-risk weight eta (ablation) | GMPQ · ImageNet ResNet18 | accuracy-complexity curve | Improved |
Closest papers by meaning
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.91)
- Efficient and Generalizable Mixed-Precision Quantization via Topological Entropy (NeurIPS 2025 · similarity 0.90)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.89)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.87)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.87)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · 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.
