Research map / Architectures and efficiency / Quantization
Improving Low-Precision Network Quantization via Bin Regularization
Tiantian Han, Dong Li, Ji Liu, Lu Tian, Yi Shan
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 bin regularization | LSQ, 2-bit · MobileNetV2 ImageNet | top-1 accuracy: 46.7 -> 50.6 | Improved |
| add bin regularization | LSQ, 2-bit · MobileNetV3-Small ImageNet | top-1 accuracy: 31.4 -> 36.3 | Improved |
| add bin regularization | LSQ, 3-bit · MobileNetV3-Small ImageNet | top-1 accuracy: 52.0 -> 56.0 | Improved |
| add bin regularization | LSQ, 2-bit · ResNet18 ImageNet | top-1 accuracy: 66.5 -> 67.2 | Improved |
| add bin regularization | LSQ, 4-bit · ResNet18 ImageNet | top-1 accuracy: 70.5 -> 70.8 | Improved |
| add bin regularization | LSQ, 4-bit · MobileNetV2 ImageNet | top-1 accuracy: 69.5 -> 70.4 | Improved |
| use KURE regularization instead of BR (ablation) | LSQ, 2-bit · MobileNetV2 ImageNet | top-1 accuracy: 46.7 -> 37.0 | Hurt |
| use L2 regularization instead of BR (ablation) | LSQ, 2-bit · MobileNetV2 ImageNet | top-1 accuracy: 46.7 -> 42.1 | Hurt |
| use L2 regularization instead of BR (ablation) | LSQ, 3-bit · MobileNetV2 ImageNet | top-1 accuracy: 65.3 -> 66.5 | Improved |
| add bin regularization from the start jointly (S1) instead of after 30 epochs of step-size learning (S2) (ablation) | LSQ + BR, 2-bit · MobileNetV2 ImageNet | top-1 accuracy: 50.6 -> 44.3 | Hurt |
| alternate step-size updates and bin regularization (S3) instead of S2 (ablation) | LSQ + BR, 2-bit · MobileNetV2 ImageNet | top-1 accuracy: 50.6 -> 43.6 | Hurt |
| vary loss weight lambda from 4 to 0.05 (ablation) | LSQ + BR, 3-bit · MobileNetV3-Small ImageNet | top-1 accuracy: 55.2 -> 54.7 | Mixed |
Closest papers by meaning
- Loss-aware Weight Quantization of Deep Networks (ICLR 2018 · similarity 0.90)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.90)
- Analysis of Quantized Models (ICLR 2019 · similarity 0.90)
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.89)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.89)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 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.
