Research map / Architectures and efficiency / Quantization
RMSMP: A Novel Deep Neural Network Quantization Framework With Row-Wise Mixed Schemes and Multiple Precisions
Sung-En Chang, Yanyu Li, Mengshu Sun, Weiwen Jiang, Sijia Liu, Yanzhi Wang, Xue Lin
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 |
|---|---|---|---|
| add 5% Fixed-W8A4 filters (row-wise multi-precision) to Fixed-W4A4 (ablation) | Fixed-W4A4 · ImageNet, ResNet-18 | top-1 accuracy (%): 68.66 -> 70.48 | Improved |
| add PoT-W4A4 filters to Fixed-W4A4 + Fixed-W8A4 (RMSMP 65:30:5) (ablation) | Fixed-W4A4 + Fixed-W8A4 · ImageNet, ResNet-18 | top-1 accuracy (%): 70.48 -> 70.73 | Improved |
| add PoT-W4A4 filters to Fixed-W4A4 + Fixed-W8A4 (RMSMP) (ablation) | Fixed-W4A4 + Fixed-W8A4 · ImageNet, ResNet-50 | top-1 accuracy (%): 76.68 -> 76.62 | No clear effect |
| add PoT-W4A4 filters to Fixed-W4A4 + Fixed-W8A4 (RMSMP) (ablation) | Fixed-W4A4 + Fixed-W8A4 · ImageNet, MobileNet-v2 | top-1 accuracy (%): 68.98 -> 69.02 | No clear effect |
| mix PoT-W4A4 with Fixed-W4A4 filters (1:1) instead of Fixed alone (ablation) | Fixed-W4A4 · ImageNet, ResNet-18 | top-1 accuracy (%): 68.66 -> 67.98 | Hurt |
| quantize first and last layers to 4-bit instead of 8-bit Fixed (ablation) | Fixed-W4A4 · ImageNet, ResNet-18 | top-1 accuracy (%): 69.72 -> 68.66 | Hurt |
| quantize first and last layers to 4-bit instead of 8-bit Fixed (ablation) | Fixed-W4A4 · FPGA XC7Z020, ResNet-18 | latency (ms): 122.6 -> 99.3 | Improved |
| use RMSMP-1 (60:35:5) instead of Fixed-only quantization | Fixed single scheme · FPGA XC7Z020, ResNet-18 | latency (ms): 122.6 -> 40.7 | Improved |
| use RMSMP-2 (65:30:5) instead of Fixed-only quantization | Fixed single scheme · FPGA XC7Z045, ResNet-18 | latency (ms): 31.4 -> 8.6 | Improved |
| use RMSMP instead of MSQ | MSQ · ImageNet, ResNet-18 | top-1 accuracy (%): 70.27 -> 70.73 | Improved |
| use RMSMP instead of MSQ-2 on XC7Z045 | MSQ-2 · FPGA XC7Z045, ResNet-18 | latency (ms): 10.1 -> 8.6 | Improved |
| quantize BERT with RMSMP rather than 32-bit | BERT · SST-2 | accuracy (%): 93.00 -> 92.87 | No clear effect |
Closest papers by meaning
- HMQ: Hardware Friendly Mixed Precision Quantization Block for CNNs (ECCV 2020 · similarity 0.89)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.88)
- Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data (AAAI 2019 · similarity 0.88)
- BSQ: Exploring Bit-Level Sparsity for Mixed-Precision Neural Network Quantization (ICLR 2021 · similarity 0.88)
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.88)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 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.
