Research map / Architectures and efficiency / Quantization
HAWQ: Hessian AWare Quantization of Neural Networks With Mixed-Precision
Zhen Dong, Zhewei Yao, 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 |
|---|---|---|---|
| add Hessian-aware mixed precision (vs direct mixed precision) (ablation) | ResNet20 W2MP A4 · CIFAR-10 | accuracy (%): 90.34 -> 92.22 | Improved |
| add Hessian-aware mixed precision (vs direct mixed precision) (ablation) | Inception-V3 W2MP A4MP · ImageNet | top-1 (%): 69.76 -> 75.52 | Improved |
| add Hessian-aware mixed precision (vs direct mixed precision) (ablation) | SqueezeNext W3MP A8 · ImageNet | top-1 (%): 65.39 -> 68.02 | Improved |
| use Hessian-based precision order instead of reversed order (ablation) | Inception-V3 · ImageNet | top-1 (%): 66.72 -> 74.26 | Improved |
| use Hessian-based precision order instead of reversed order (ablation) | Inception-V3 · ImageNet | epochs to converge: 50 -> 30 | Improved |
| use descending Omega block-wise fine-tuning order instead of reverse order (ablation) | Inception-V3 · ImageNet | epochs to converge for a block: 50 -> 25 | Improved |
| use HAWQ instead of LQ-Nets 3-bit | ResNet20 · CIFAR-10 | accuracy (%): 91.60 -> 92.22 | Improved |
| use HAWQ instead of HAQ on ResNet50 | ResNet50 · ImageNet | top-1 (%): 75.30 -> 75.48 | Improved |
| use HAWQ instead of RVQuant | Inception-V3 · ImageNet | top-1 (%): 74.14 -> 75.52 | Improved |
| use HAWQ instead of DNAS | ResNet20 · CIFAR-10 | activation compression: 1.00x -> 8.00x | Improved |
Closest papers by meaning
- HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks (NeurIPS 2020 · similarity 0.95)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.91)
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.91)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.88)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.87)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · 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.
