Research map / Architectures and efficiency / Quantization
QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization
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 |
|---|---|---|---|
| Randomly drop activation quantization during PTQ reconstruction (ablation) | ImageNet models quantized to W2A2 · ImageNet, ResNet-18 | accuracy: 46.64 -> 51.14 | Helped |
| Randomly drop activation quantization during PTQ reconstruction (ablation) | ImageNet models quantized to W2A2 · ImageNet, RegNet-3.2GF | accuracy: 39.76 -> 52.36 | Helped |
| Randomly drop activation quantization during PTQ reconstruction (ablation) | ImageNet models quantized to W2A4 · ImageNet, MNasNet-2.0 | accuracy: 60.00 -> 62.36 | Helped |
| Include activation quantization during weight reconstruction instead of excluding it (ablation) | PTQ reconstruction · ImageNet W2A2, ResNet-18 | accuracy: 18.88 -> 45.74 | Helped |
| Omit activation quantization in the current tuning block rather than quantizing all activations (ablation) | PTQ reconstruction · ImageNet W2A2, ResNet-18 | accuracy: 45.74 -> 48.07 | Helped |
| Use full activation quantization during reconstruction instead of randomly dropping it (ablation) | PTQ reconstruction · ImageNet W2A2, ResNet-18 | accuracy: 51.14 -> 45.77 | Hurt |
| Set activation quantization dropping probability to 0.5 (ablation) | QDrop · ImageNet, MobileNetV2 and RegNet-600MF | accuracy | Helped |
| Randomly drop activation quantization (ablation) | QDrop on BERT · GLUE QNLI (E8W4A4) | accuracy: 68.05 -> 76.75 | Helped |
| Randomly drop activation quantization (ablation) | QDrop on BERT · SQuAD1.1 (E8W4A4) | F1: 75.97 -> 77.26 | Helped |
| Randomly drop activation quantization (ablation) | QDrop with cross-domain calibration · ImageNet, CIFAR100 calibration, W4A4 | accuracy: 46.83 -> 52.88 | Helped |
Closest papers by meaning
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.87)
- Enhancing Post-training Quantization Calibration through Contrastive Learning (CVPR 2024 · similarity 0.87)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.87)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.87)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.87)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 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.