Research map / Architectures and efficiency / Quantization
Nonuniform-to-Uniform Quantization: Towards Accurate Quantization via Generalized Straight-Through Estimation
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 |
|---|---|---|---|
| use N2UQ instead of LCQ | ResNet-18 · ImageNet, W2/A2 | top-1 accuracy: 68.9 -> 69.4 | Helped |
| use N2UQ instead of LCQ | ResNet-34 · ImageNet, W2/A2 | top-1 accuracy: 72.7 -> 73.3 | Helped |
| use N2UQ instead of LCQ | ResNet-50 · ImageNet, W2/A2 | top-1 accuracy: 75.1 -> 75.8 | Helped |
| use N2UQ instead of PROFIT | MobileNet-V2 · ImageNet | top-1 accuracy: 71.6 -> 72.1 | Helped |
| add threshold-learning activation quantization with G-STE (ablation) | 2-bit ResNet-18 baseline · ImageNet | top-1 accuracy: 65.9 -> 68.9 | Helped |
| add entropy-preserving weight regularization (ablation) | 2-bit ResNet-18 baseline · ImageNet | top-1 accuracy: 65.9 -> 67.8 | Helped |
| add both proposed components (ablation) | 2-bit ResNet-18 baseline · ImageNet | top-1 accuracy: 65.9 -> 69.7 | Helped |
| replace entropy-preserving regularization with Weight Norm (ablation) | 2-bit ResNet-18 with G-STE threshold learner · ImageNet | top-1 accuracy: 69.7 -> 67.4 | Hurt |
| replace entropy-preserving regularization with learnable scaling factor (ablation) | 2-bit ResNet-18 with G-STE threshold learner · ImageNet | top-1 accuracy: 69.7 -> 68.6 | Hurt |
| keep 1×1 downsampling layers full-precision (ablation) | 2-bit ResNet-50 N2UQ · ImageNet | top-1 accuracy: 75.8 -> 76.4 | Helped |
| keep 1×1 downsampling layers full-precision (ablation) | 4-bit ResNet-50 N2UQ · ImageNet | top-1 accuracy: 78.0 -> 78.0 | No clear effect |
Closest papers by meaning
- Mr.BiQ: Post-Training Non-Uniform Quantization Based on Minimizing the Reconstruction Error (CVPR 2022 · similarity 0.89)
- Additive Powers-of-Two Quantization: An Efficient Non-uniform Discretization for Neural Networks (ICLR 2020 · similarity 0.89)
- Learnable Companding Quantization for Accurate Low-Bit Neural Networks (CVPR 2021 · similarity 0.89)
- Universally Quantized Neural Compression (NeurIPS 2020 · similarity 0.89)
- PowerQuant: Automorphism Search for Non-Uniform Quantization (ICLR 2023 · similarity 0.88)
- Quantization without Tears (CVPR 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.