Research map / Architectures and efficiency / Quantization
Channel Balancing for Accurate Quantization of Winograd Convolutions
Check 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 |
|---|---|---|---|
| Balance Winograd-domain input and filter channels | 8-bit static tile-scale PTQ ResNet models using F(6, 3) · ImageNet | accuracy: 53.56% -> 60.59%; 62.34% -> 66.52%; 73.75% -> 74.47% | Helped |
| Balance Winograd-domain input and filter channels | ResNet-20 with F(4, 3), 8-bit static tile-scale PTQ · CIFAR-10 | accuracy: 89.93 -> 90.31 | Helped |
| Balance Winograd-domain input and filter channels | ESPCNN with F(4, 3), 8-bit static tile-scale PTQ · Set5 and Set14, 3× super-resolution | PSNR to ground truth: 29.41 dB -> 30.24 dB; 26.33 dB -> 26.76 dB | Helped |
| Balance channels and train the balancing coefficients during QAT | ResNet-20 with F(6, 3), 8-bit static tile-scale QAT · CIFAR-10 | accuracy: 48.77 -> 90.57 | Helped |
| Balance channels and train the balancing coefficients during QAT | ResNet-20 with F(4, 3), 4-bit static tile-scale QAT · CIFAR-10 | accuracy: 33.79 -> 85.33 | Helped |
| Use static rather than dynamic quantization (ablation) | ResNet-20 with F(6, 3), 8-bit tile scales · CIFAR-10 | accuracy: 81.29 -> 90.90 (QW); 81.44 -> 91.18 (BQW) | Mixed |
| Use tile rather than scalar quantization scales (ablation) | ResNet-20 with 8-bit static quantization · CIFAR-10 | accuracy: 16.22 -> 90.31 | Helped |
| Use the F(6, 3) rather than F(4, 3) Winograd algorithm (ablation) | ResNet-20, 8-bit static tile-scale QAT · CIFAR-10 | accuracy: 91.75 -> 90.57 | Hurt |
Closest papers by meaning
- Towards Efficient and Accurate Winograd Convolution via Full Quantization (NeurIPS 2023 · similarity 0.92)
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.91)
- Efficient Sparse-Winograd Convolutional Neural Networks (ICLR 2018 · similarity 0.87)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.83)
- SFC: Achieve Accurate Fast Convolution under Low-precision Arithmetic (ICML 2024 · similarity 0.82)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.82)
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.