Research map / Architectures and efficiency / Quantization
Accurate Post Training Quantization With Small Calibration Sets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Layer-wise AdaQuant (joint weight and quantization parameter optimization) instead of rounding-only or range-only methods | ResNet-50 INT4 post-training quantization · ImageNet | top-1 accuracy | Helped |
| Sequential AdaQuant (errors from previous quantized layers accounted for) | 4-bit ResNet models · ImageNet | top-1 accuracy: ResNet-50 73.7% -> 75.1% | Helped |
| Block-wise sequential AdaQuant instead of layer-wise | ResNet-50 INT4 · ImageNet | top-1 accuracy: 75.1% -> 75.9% | Helped |
| Integer programming bit allocation vs greedy-compression bit allocation (ablation) | ResNet-50 and MobileNet-V2 mixed precision · ImageNet ablation (Fig. 4) | top-1 accuracy at compression 0.15-0.18: about 10 points over greedy-c | Helped |
| Para-Normalization BN statistics re-tuning in light pipeline (ablation) | Light pipeline on ResNet-50 at challenging compression · ImageNet ablation (Fig. 4) | top-1 accuracy: 73.7% -> 75.37% | Helped |
| Layer-wise AdaQuant on 8-bit MobileNet-V2 (vs min-max calibration) | MobileNet-V2 INT8 · ImageNet | top-1 accuracy: 70.9% -> 73.03% | Helped |
| Layer-wise AdaQuant on INT8 BERT-base | BERT-base SQuAD1.1 · SQuAD1.1 | F1: 87.83 -> 88.35 | Helped |
| Sequential AdaQuant vs layer-wise AdaQuant on INT8 MobileNet-V2 | MobileNet-V2 INT8 · ImageNet | top-1 accuracy: 73.03% vs 72.94% | No clear effect |
| AdaQuant with a single calibration image (no overfitting observed as calibration set size grows) (ablation) | ResNet-50 INT4 AdaQuant · ImageNet calibration set size | top-1 accuracy | No clear effect |
Closest papers by meaning
- Enhancing Post-training Quantization Calibration through Contrastive Learning (CVPR 2024 · similarity 0.90)
- Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (NeurIPS 2026 · similarity 0.90)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.89)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.88)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.88)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (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.