Research map / Architectures and efficiency / Quantization
Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization
Weiying Xie, Zihan Meng, Jitao Ma, Wenjin Guo, Haowei Li, Haonan Qin, Leyuan Fang, Yunsong Li
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 early-stage oscillation encouragement and decoupled thresholds/levels (expand solution space) (ablation) | LSQ, ResNet-50 W2/A2 · ImageNet val | top-1 accuracy (%): 73.7 -> 75.2 | Improved |
| add delayed weight oscillation suppression (ablation) | LSQ, ResNet-50 W2/A2 · ImageNet val | top-1 accuracy (%): 73.7 -> 75.6 | Improved |
| combine expanded solution space and delayed suppression (ablation) | LSQ, ResNet-50 W2/A2 · ImageNet val | top-1 accuracy (%): 73.7 -> 76.3 | Improved |
| allow oscillation early instead of suppressing it throughout training (ablation) | LSQ toy example · toy 2-bit problem | minimum loss: 26.2 -> 5.6 | Improved |
| suppress oscillation only in final stage versus throughout (ablation) | LSQ toy example · toy 2-bit problem | minimum loss: 26.2 -> 16.7 | Improved |
| add AOQ to LSQ on 2-bit ResNet-18 | LSQ · ImageNet val ResNet-18 W2/A2 | top-1 accuracy (%): 67.6 -> 69.7 | Improved |
| add AOQ to N2UQ on 4-bit ResNet-18 | N2UQ · ImageNet val ResNet-18 W4/A4 | top-1 accuracy (%): 72.9 -> 72.7 | Hurt |
| add AOQ to N2UQ on ResNet-34/50 at 3/4 bits | N2UQ · ImageNet val ResNet-34 W4/A4 | top-1 accuracy (%): 76.0 -> 76.0 | No clear effect |
| use AOQ on DeiT-T at 3-bit versus OFQ | LSQ+AOQ · ImageNet val DeiT-T W3/A3 | top-1 accuracy (%): 72.7 -> 72.3 | Hurt |
| use AOQ instead of OFQ on 2-bit DeiT-T | LSQ+AOQ · ImageNet val DeiT-T W2/A2 | top-1 accuracy (%): 64.3 -> 66.5 | Improved |
| use AOQ instead of oscillation suppression throughout training on MobileNet-V2 | LSQ+AOQ · ImageNet val MobileNet-V2 W3/A3 | top-1 accuracy (%): 67.8 -> 69.1 | Improved |
Closest papers by meaning
- Overcoming Oscillations in Quantization-Aware Training (ICML 2022 · similarity 0.92)
- Oscillation-free Quantization for Low-bit Vision Transformers (ICML 2023 · similarity 0.92)
- Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective (CVPR 2023 · similarity 0.89)
- Once Quantization-Aware Training: High Performance Extremely Low-Bit Architecture Search (ICCV 2021 · similarity 0.88)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.86)
- A2Q+: Improving Accumulator-Aware Weight Quantization (ICML 2024 · similarity 0.86)
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.
