Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Fine-Grained Data Distribution Alignment for Post-Training Quantization
Yunshan Zhong, Mingbao Lin, Mengzhao Chen, Ke Li, Yunhang Shen, Fei Chao, Yongjian Wu, Rongrong Ji
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 real calibration dataset to synthetic data (ablation) | GDFQ · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 60.60 -> 65.64 | Improved |
| add fine-grained BNS alignment (BNS-centralized and BNS-distorted losses) (ablation) | GDFQ + C · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 65.64 -> 68.88 | Improved |
| use predicted labels instead of real labels for calibration images (ablation) | FDDA · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 68.88 -> 68.68 | No clear effect |
| FDDA versus best zero-shot method with calibration data (DSG) | FDDA · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 66.83 -> 68.88 | Improved |
| FDDA versus BRECQ at W4A4 F4L4 on MobileNetV1 | FDDA · ImageNet val (W4A4, F4L4) | top-1 accuracy (%): 57.11 -> 63.75 | Improved |
| FDDA versus BRECQ at W4A4 F8L8 on MobileNetV2 | FDDA · ImageNet val (W4A4, F8L8) | top-1 accuracy (%): 66.57 -> 69.32 | Improved |
| FDDA versus BRECQ at W5A5 F8L8 on ResNet-18 | FDDA · ImageNet val (W5A5, F8L8) | top-1 accuracy (%): 70.60 -> 70.86 | Improved |
| reduce available calibration classes to 700 (ablation) | FDDA · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 68.09 | Hurt |
| FDDA with only 700 classes versus BRECQ | FDDA · ImageNet val (ResNet-18, W4A4 all layers) | top-1 accuracy (%): 67.94 -> 68.09 | Improved |
Closest papers by meaning
- Accurate Post Training Quantization With Small Calibration Sets (ICML 2021 · similarity 0.89)
- Post Training Quantization for Efficient Dataset Condensation (AAAI 2026 · similarity 0.88)
- Sharpness-Aware Data Generation for Zero-shot Quantization (ICML 2024 · similarity 0.86)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.86)
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.85)
- Enhancing Post-training Quantization Calibration through Contrastive Learning (CVPR 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.
