Research map / Architectures and efficiency / Quantization
Post-Training Piecewise Linear Quantization for Deep Neural Networks
Jun Fang, Ali Shafiee, Hamzah Abdel-Aziz, David Thorsley, Georgios Georgiadis, Joseph H. Hassoun
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 piecewise linear instead of uniform quantization (4-bit weights) | Inception-v3 · ImageNet val | top-1 accuracy (%): 44.28 -> 75.72 | Improved |
| use piecewise linear instead of uniform quantization (4-bit weights) | ResNet-50 · ImageNet val | top-1 accuracy (%): 65.48 -> 74.28 | Improved |
| use piecewise linear instead of uniform quantization (4-bit weights) | MobileNet-v2 · ImageNet val | top-1 accuracy (%): 11.37 -> 54.34 | Improved |
| use piecewise linear instead of uniform quantization (8-bit weights) | Inception-v3 · ImageNet val | top-1 accuracy (%): 77.53 -> 77.52 | No clear effect |
| add bias correction to 4-bit PWLQ (ablation) | MobileNet-v2 · ImageNet val | top-1 accuracy (%): 54.34 -> 69.22 | Improved |
| add bias correction to 4-bit PWLQ (ablation) | Inception-v3 · ImageNet val | top-1 accuracy (%): 75.72 -> 76.45 | Improved |
| use two breakpoints instead of one (3-bit weights) (ablation) | MobileNet-v2 · ImageNet val | top-1 accuracy (%): 16.77 -> 36.44 | Improved |
| use three breakpoints instead of one (3-bit weights) (ablation) | MobileNet-v2 · ImageNet val | top-1 accuracy (%): 16.77 -> 55.17 | Improved |
| use two breakpoints instead of one (4-bit weights) (ablation) | ResNet-50 · ImageNet val | top-1 accuracy (%): 74.28 -> 75.24 | Improved |
| perturb optimal breakpoints by 5% (ablation) | 4-bit Inception-v3 · ImageNet val | top-1 accuracy (%): 75.72 -> 74.05 | Hurt |
| use PWLQ instead of uniform quantization (4-bit weights) | DeepLab-v3+ · Pascal VOC segmentation | mIoU (%): 50.05 -> 67.66 | Improved |
| use PWLQ instead of uniform quantization (4-bit weights) | SSD-Lite · Pascal VOC detection | mAP (%): 64.79 -> 68.32 | Improved |
Closest papers by meaning
- Non-uniform Step Size Quantization for Accurate Post-Training Quantization (ECCV 2022 · similarity 0.89)
- Optimization-Based Post-Training Quantization With Bit-Split and Stitching (TPAMI 2022 · similarity 0.89)
- Pushing the Limit of Post-Training Quantization (TPAMI 2025 · similarity 0.89)
- Up or Down? Adaptive Rounding for Post-Training Quantization (ICML 2020 · similarity 0.88)
- FlexRound: Learnable Rounding based on Element-wise Division for Post-Training Quantization (ICML 2023 · similarity 0.88)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.88)
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.
