Research map / Architectures and efficiency / Quantization
HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks
Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W. Mahoney, Kurt Keutzer
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 average Hessian trace rather than parameter-perturbation L2 alone to rank quantization sensitivity (ablation) | HAWQ-V2 · ImageNet, SqueezeNext | Top-1 accuracy: 67.83 -> 68.68 | Improved |
| choose the candidate precision setting with minimum total perturbation (ablation) | HAWQ-V2 · ImageNet, SqueezeNext | Top-1 accuracy: 67.46 -> 68.68 | Improved |
| replace HAWQ's top-eigenvalue sensitivity with trace-weighted automatic precision selection | HAWQ · ImageNet, ResNet50 | Top-1 accuracy: 75.48 -> 75.92 | Improved |
| replace manual mixed-precision selection with HAWQ-V2 automatic selection | HAWQ · ImageNet, InceptionV3 | Top-1 accuracy: 75.52 -> 75.98 | Improved |
| use trace-weighted automatic mixed precision instead of HAWQ | HAWQ · ImageNet, SqueezeNext | Top-1 accuracy: 68.02 -> 68.68 | Improved |
| use HAWQ-V2 rather than FQN for quantization | RetinaNet-ResNet50 · Microsoft COCO 2017 | mAP: 32.5 -> 34.1 | Improved |
| use HAWQ-V2 rather than direct uniform quantization | RetinaNet-ResNet50 · Microsoft COCO 2017 | mAP: 31.5 -> 34.1 | Improved |
| use trace-weighted HAWQ-V2 rather than HAWQ | RetinaNet-ResNet50 · Microsoft COCO 2017 | mAP: 33.5 -> 34.1 | Improved |
| use mixed-precision rather than uniform 4-bit activations | HAWQ-V2 · Microsoft COCO 2017 | mAP: 34.1 -> 34.4 | Improved |
| increase activation precision from 4-bit to 6-bit | HAWQ-V2 · Microsoft COCO 2017 | mAP: 34.1 -> 34.8 | Improved |
| use mixed-precision quantization rather than direct quantization | RetinaNet-ResNet50 · Microsoft COCO 2017 | mAP: 31.5 -> 34.1 | Improved |
Later papers that compared with this method
4 later accepted papers report HAWQv2 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · better on 1 of 2 shared results)
- EMQ: Evolving Training-free Proxies for Automated Mixed Precision Quantization (ICCV 2023 · better on 1 of 1 shared result)
- Mixed-Precision Neural Network Quantization via Learned Layer-Wise Importance (ECCV 2022 · better on 1 of 1 shared result)
- Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization (ICCV 2021 · better on 0 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- HAWQ: Hessian AWare Quantization of Neural Networks With Mixed-Precision (ICCV 2019 · similarity 0.95)
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.90)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.89)
- Network Quantization With Element-Wise Gradient Scaling (CVPR 2021 · similarity 0.88)
- WinQ: Accelerating Quantization-Aware Training of Language Models Around Saddle Points (ICML 2026 · similarity 0.87)
- HMQ: Hardware Friendly Mixed Precision Quantization Block for CNNs (ECCV 2020 · 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.
