Research map / Architectures and efficiency / CNN design
Efficient Residue Number System Based Winograd Convolution
Zhi-Gang Liu, Matthew Mattina
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 16-bit RNS(4001,4331) instead of 8-bit RNS(251,241,239) (ablation) | F(14x14,3x3) Winograd · VGG16 INT8 on Arm Cortex-A73 | average speed-up: 2.02 -> 2.20 | Improved |
| use 2-residue 16-bit RNS instead of 3-residue 8-bit RNS (ablation) | F(10x10,3x3) Winograd · Arithmetic complexity reduction | arithmetic complexity reduction: 2.08 -> 3.13 | Improved |
| use 2-residue RNS(4001,4331) instead of 3-residue RNS for large tiles (ablation) | F(12x12,5x5) Winograd · Arithmetic complexity reduction | arithmetic complexity reduction: 4.69 -> 7.03 | Improved |
| increase Winograd tile from 2x2 to 14x14 output (ablation) | Winograd over RNS(251,241,239) · Arithmetic complexity reduction (3x3) | arithmetic complexity reduction: 0.75 -> 2.30 | Improved |
| use 5x5 filters instead of 3x3 filters in RNS Winograd (ablation) | F(12x12,k) over RNS(251,241,239) · Arithmetic complexity reduction | arithmetic complexity reduction: 2.20 -> 4.69 | Improved |
| move from conv1_2 to deeper VGG16 layers (ablation) | F(14x14,3x3) RNS(251,241,239) · VGG16 layers | speed-up: 1.86 -> 2.22 | Improved |
| apply RNS Winograd to 8-bit CNNs without precision scaling | RNS Winograd · Quantized CNN accuracy | top-1 accuracy | No clear effect |
| use larger Winograd tile in standard integer Winograd (no RNS) (ablation) | Integer Winograd · Transform data width | transform data width (bits): 12 -> 18 | Hurt |
Closest papers by meaning
- Towards Efficient and Accurate Winograd Convolution via Full Quantization (NeurIPS 2023 · similarity 0.88)
- Efficient Sparse-Winograd Convolutional Neural Networks (ICLR 2018 · similarity 0.88)
- DWM: A Decomposable Winograd Method for Convolution Acceleration (AAAI 2020 · similarity 0.87)
- SFC: Achieve Accurate Fast Convolution under Low-precision Arithmetic (ICML 2024 · similarity 0.84)
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.83)
- A Fast Convoluted Story: Scaling Probabilistic Inference for Integer Arithmetics (NeurIPS 2024 · similarity 0.80)
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.
