Research map / Architectures and efficiency / Quantization
Improving Neural Network Efficiency via Post-Training Quantization With Adaptive Floating-Point
Fangxin Liu, Wenbo Zhao, Zhezhi He, Yanzhi Wang, Zongwu Wang, Changzhi Dai, Xiaoyao Liang, Li Jiang
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 dynamic instead of static activation configuration (4.8-bit) (ablation) | AFP ResNet-50 · ImageNet val | top-1 accuracy (%): 76.00 -> 76.09 | Improved |
| use dynamic instead of static activation configuration (3.9-bit) (ablation) | AFP ResNet-50 · ImageNet val | top-1 accuracy (%): 75.11 -> 75.27 | Improved |
| use dynamic instead of static activation configuration (5.7-bit) (ablation) | AFP MobileNet-v2 · ImageNet val | top-1 accuracy (%): 71.47 -> 71.59 | Improved |
| use dynamic instead of static activation configuration (4.8-bit) (ablation) | AFP MobileNet-v2 · ImageNet val | top-1 accuracy (%): 70.91 -> 71.11 | Improved |
| reduce AFP bit-width from 4.8 to 3.9 bits (dynamic) (ablation) | AFP ResNet-50 · ImageNet val | top-1 accuracy (%): 76.09 -> 75.27 | Hurt |
| reduce AFP bit-width from 5.7 to 4.8 bits (dynamic) (ablation) | AFP MobileNet-v2 · ImageNet val | top-1 accuracy (%): 71.59 -> 71.11 | Hurt |
| search AFP configuration with Bayesian optimization instead of enumeration (ablation) | AFP bit-width search · MobileNet-v2 and ResNet-50 | search iterations: 35 -> 18 iterations, 48.8% less time | Improved |
| use AFP over fixed 5-bit-style integer baselines at same bit-width (vs INT8) | AFP dynamic vs INT8 · ImageNet val ResNet-50 | top-1 accuracy (%): 74.9 -> 76.09 | Improved |
| replace conventional quantizer/MAC with AFP shift-and-add hardware | AFP-encoded models · 90 nm synthesis | energy: 9.3x for MAC, 13.2x for quantizer | Improved |
| increase calibration batch size up to 32 (ablation) | AFP static quantization · ImageNet val | top-1 accuracy: accuracy stabilizes at batch size 32 | Improved |
| raise lambda from 0.25 to 0.50 in search objective (ablation) | AFP search · MobileNet-v2 ImageNet | chosen bit-width Nq: Nq 7 -> 5 | Improved |
Closest papers by meaning
- Be Like Water: Adaptive Floating Point for Machine Learning (ICML 2022 · similarity 0.92)
- Adaptive Quantization of Neural Networks (ICLR 2018 · similarity 0.89)
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.87)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.86)
- MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization (NeurIPS 2024 · similarity 0.85)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.85)
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.
