Research map / Architectures and efficiency / Quantization
Mixed-Precision Neural Network Quantization via Learned Layer-Wise Importance
Chen Tang, Kai Ouyang, Zhi Wang, Yifei Zhu, Wen Ji, Yaowei Wang, Wenwu Zhu
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 |
|---|---|---|---|
| assign more bits to layers with larger learned scale factors (not reversed) (ablation) | MobileNetV1 W4MP/A4MP · ImageNet 9.68G BitOps | top-1 accuracy (%): 65.25 -> 71.84 | Improved |
| use importance-guided assignment instead of reversed at 3-bit level versus reversed 4-bit (ablation) | MobileNetV1 · ImageNet | top-1 accuracy (%): 65.25 -> 69.48 | Improved |
| initialize from a higher-accuracy FP model (70.5 vs 69.6) | Ours ResNet-18 W3MP/A3MP · ImageNet | top-1 accuracy (%): 69.0 -> 69.7 | Improved |
| initialize from a higher-accuracy FP model (70.5 vs 69.6) | Ours ResNet-18 W4MP/A4MP · ImageNet | top-1 accuracy (%): 70.1 -> 70.8 | Improved |
| tighten constraint from 3W3A to 2.5W3A (19.81G BitOps) | Ours ResNet-18 · ImageNet | top-1 accuracy (%): 69.0 -> 68.7 | Hurt |
| use learned-importance MPQ instead of DNAS at 3-bit | ResNet-18 · ImageNet | top-1 accuracy (%): 68.7 -> 69.7 | Improved |
| use learned-importance MPQ instead of HAWQv2 (ILP search) | ResNet-50 · ImageNet | top-1 accuracy (%): 75.8 -> 76.9 | Improved |
| use learned-importance MPQ instead of HAQ (3MP weights) | MobileNetV1 weight-only · ImageNet | top-1 accuracy (%): 67.66 -> 71.57 | Improved |
| use learned-importance MPQ instead of HAQ (4MP weights) | MobileNetV1 weight-only · ImageNet | top-1 accuracy (%): 71.74 -> 72.60 | Improved |
| use learned-importance MPQ instead of SPOS at 4-bit | ResNet-18 · ImageNet | top-1 accuracy (%): 70.5 -> 70.8 | Improved |
| use ILP search on learned importance instead of AutoQ search | ResNet-50 search time · policy search cost | search time: 1000z GPU-hours -> 330z x speedup | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HAWQv2 — HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.93)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.91)
- Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization (ICCV 2021 · similarity 0.90)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.90)
- BSQ: Exploring Bit-Level Sparsity for Mixed-Precision Neural Network Quantization (ICLR 2021 · similarity 0.90)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.88)
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.
