Research map / Architectures and efficiency / Quantization
Solving Oscillation Problem in Post-Training Quantization Through a Theoretical Perspective
Yuexiao Ma, Huixia Li, Xiawu Zheng, Xuefeng Xiao, Rui Wang, Shilei Wen, Xin Pan, Fei Chao, Rongrong Ji
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 MRECG to BRECQ | BRECQ · ImageNet val, ResNet-18 W/A 2/4 | Top-1 accuracy (%): 63.71 -> 65.61 (+1.9) | Improved |
| Add MRECG to BRECQ | BRECQ · ImageNet val, MobileNetV2 ×0.5 W/A 2/4 | Top-1 accuracy (%): 34.55 -> 41.16 (+6.61) | Improved |
| Add MRECG to QDROP | QDROP · ImageNet val, Reg600M W/A 2/4 | Top-1 accuracy (%): 63.10 -> 65.16 (+2.06) | Improved |
| Add mixed reconstruction granularity (ablation) | BRECQ on MobileNetV2 ×0.5 W/A 4/4 · ImageNet val | Top-1 accuracy (%): 50.83 -> 51.91 | Improved |
| Expand calibration batch size (ablation) | BRECQ on MobileNetV2 ×0.5 W/A 4/4 · ImageNet val | Top-1 accuracy (%): 50.83 -> 51.57 | Improved |
| Combine MRECG with expanded calibration batch size (ablation) | BRECQ on MobileNetV2 ×0.5 W/A 4/4 · ImageNet val | Top-1 accuracy (%): 50.83 -> 52.25 | Improved |
| Use reconstruction loss rather than ModCap to choose module combinations (ablation) | MRECG · ImageNet val, ResNet-18 W/A 3/3 | Top-1 accuracy (%): 66.07 -> 66.30 | Improved |
| Use reconstruction loss rather than ModCap to choose module combinations (ablation) | MRECG · ImageNet val, MobileNetV2 ×0.5 W/A 3/3 | Top-1 accuracy (%): 38.02 -> 38.43 | Improved |
| Increase calibration batch size (ablation) | PTQ reconstruction · ImageNet val, MobileNetV2 ×0.5 W/A 4/4 | Top-1 accuracy (%) | Mixed |
| Jointly optimize adjacent modules with large capacity differences (ablation) | PTQ reconstruction · ImageNet classification | Reconstruction loss oscillation | 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.
- QDROP — QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization (ICLR 2022) · better on 4 of 6 shared results
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
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.89)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.88)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.87)
- Enhancing Post-training Quantization Calibration through Contrastive Learning (CVPR 2024 · similarity 0.87)
- Predicting Quantization Price for Selecting PTQ Configurations Before Deployment (NeurIPS 2026 · similarity 0.87)
- Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization (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.
