Research map / Architectures and efficiency / Quantization
Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization
Weihan Chen, Peisong Wang, Jian Cheng
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 |
|---|---|---|---|
| use second-order term instead of first-order only for loss perturbation (ablation) | mixed-precision bit-width assignment · ImageNet, ResNet-18/50, MobileNet-V2 | accuracy | Improved |
| use Hessian-based loss perturbation instead of Hessian-free 1/2 dw^T dw (ablation) | bit-width assignment · ImageNet, deep and lightweight nets | accuracy | Improved |
| add first-order term to second-order term (ablation) | loss perturbation criterion · ImageNet | accuracy | Hurt |
| use original greedy priority instead of reversed criterion (ablation) | MCKP greedy search · ImageNet | accuracy | Improved |
| use original greedy priority instead of random criterion (ablation) | MCKP greedy search · ImageNet | accuracy | Improved |
| use mixed precision (Ours) instead of LQ-Nets uniform 4-bit weights | ResNet-18 · ImageNet | Top-1 drop: -0.30 -> -0.26 | Improved |
| use our method instead of HAWQ | ResNet-50 W2MP A4MP · ImageNet | Top-1 drop: -1.91 -> -0.85 | Improved |
| use our method instead of HAQ | MobileNet-V2 (13.99x vs 14.07x) · ImageNet | Top-1 accuracy: 66.75 -> 68.52 | Improved |
| use only 1024 sampled images for loss perturbation | bit-width assignment · ImageNet | loss perturbation convergence | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- HAWQ-V2 — HAWQ-V2: Hessian Aware trace-Weighted Quantization of Neural Networks (NeurIPS 2020) · better on 0 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
- A Review of State-of-the-art Mixed-Precision Neural Network Frameworks (TPAMI 2024 · similarity 0.92)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.91)
- FracBits: Mixed Precision Quantization via Fractional Bit-Widths (AAAI 2021 · similarity 0.91)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.89)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.89)
- 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.
