Research map / Architectures and efficiency / Quantization
Rethinking Asymmetric Quantization: Hidden Symmetry in Vision Model Weights
Masafumi Mori, Shinya Gongyo, Mitsuru Ambai
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 |
|---|---|---|---|
| replace zero-point AsymQ weights with dense SymQ + 98% sparse SymQ outliers | RepQ-ViT · ImageNet ViT-B W3/A4 | top-1 accuracy (%): 26.98 -> 75.24 | Improved |
| replace AsymQ and DUQ with DASQ | ERQ · ImageNet ViT-B W3/A4 | top-1 accuracy (%): 72.37 -> 78.44 | Improved |
| replace AsymQ with DASQ | RepQ-ViT · ImageNet ViT-B W4/A4 | top-1 accuracy (%): 68.48 -> 78.76 | Improved |
| replace AsymQ and DUQ with DASQ | ERQ · ImageNet ViT-B W4/A4 | top-1 accuracy (%): 78.67 -> 80.04 | Improved |
| replace AsymQ with DASQ | QDrop · ImageNet ResNet-50 W2/A4 | top-1 accuracy (%): 70.08 -> 72.63 | Improved |
| replace AsymQ with DASQ | RepQ-ViT · COCO Mask R-CNN Swin-T | box AP: 36.1 -> 41.5 | Improved |
| replace AsymQ with DASQ | RepQ-ViT · COCO Mask R-CNN Swin-S | box AP: 44.2 -> 44.7 | Improved |
| replace AsymQ with DASQ on FPGA | AsymQ W4/A4 · FPGA ViT-S ImageNet | top-1 accuracy (%): up to +6.7% | Improved |
| add 4-bit sparse matrix (98% sparsity) (ablation) | RepQ-ViT SymQ vs AsymQ dense · ViT-S W4/A4 | top-1 accuracy (%): SymQ/AsymQ gap narrows to 0.19% at p=0.95 | Improved |
| use AsymQ instead of SymQ without sparse matrix (ablation) | RepQ-ViT · ViT-S W5/A4 | top-1 accuracy (%): 71.77 -> 71.98 | No clear effect |
| use AsymQ instead of SymQ without sparse matrix (ablation) | RepQ-ViT · ViT-S W4/A4 | top-1 accuracy (%): 61.56 -> 64.62 | Improved |
| increase sparse bit width from 4 to 8 (ablation) | DASQ · FPGA ViT-S | top-1 accuracy (%): slight LUT/power increase | Improved |
Closest papers by meaning
- VLM-PTQ: Efficient Post-Training Quantization for Large Vision-Language Models (CVPR 2026 · similarity 0.88)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.88)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.87)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.87)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.87)
- GPTAQ: Efficient Finetuning-Free Quantization for Asymmetric Calibration (ICML 2025 · 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.
