Research map / Architectures and efficiency / Quantization
Towards Efficient and Accurate Winograd Convolution via Full Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 BQW with PAW at 6-bit F(6,3) | Winograd PTQ · CIFAR-10 ResNet-20 | accuracy: 39.75 -> 85.29 | Helped |
| replace BQW with PAW at 6-bit F(4,3) | Winograd PTQ · CIFAR-10 ResNet-20 | accuracy: 79.36 -> 90.25 | Helped |
| replace BQW with PAW at 4-bit F(4,3) | Winograd PTQ · CIFAR-10 ResNet-20 | accuracy: 17.33 -> 62.50 | Helped |
| replace BQW with PAW at 8-bit F(4,3) | Winograd PTQ · CIFAR-10 ResNet-20 | accuracy: 90.31 -> 92.02 | Helped |
| replace BQW with PAW at 4-bit F(6,3) | Winograd PTQ · CIFAR-10 ResNet-20 | accuracy: 10.00 -> 21.71 | Helped |
| add PAW to strong baseline, 8-bit F(6,3) | strong baseline · ImageNet ResNet-18 | top-1 accuracy: 65.84 -> 68.36 | Helped |
| quantize O with per-tensor scale in 8-bit Winograd (ablation) | ResNet-18 Winograd · ImageNet | accuracy: 69.76 -> 0.20 | Hurt |
| quantize O with FSQ instead of per-tensor | ResNet-18 Winograd · ImageNet | accuracy: 0.216 -> 67.866 | Helped |
| quantize O per-pixel instead of FSQ | ResNet-18 Winograd · ImageNet | accuracy: 67.866 -> 68.740 | No clear effect |
| add FSQ to PAW for full quantization, 8-bit F(4,3) (ablation) | PAW · ImageNet ResNet-18 | top-1 accuracy: 0.21 -> 68.16 | Helped |
| add FSQ to PAW for full quantization, 6-bit F(6,3) (ablation) | PAW · ImageNet ResNet-50 | top-1 accuracy: 5.97 -> 73.84 | Helped |
| optimize A, B and G jointly instead of none (ablation) | PAW · ImageNet ResNet-18 | accuracy: 68.04 -> 69.26 | Helped |
Closest papers by meaning
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.92)
- SFC: Achieve Accurate Fast Convolution under Low-precision Arithmetic (ICML 2024 · similarity 0.87)
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.85)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.85)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.85)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.84)
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.