Research map / Architectures and efficiency / Quantization
Distribution Adaptive INT8 Quantization for Training CNNs
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 |
|---|---|---|---|
| Use channel-wise vectorized gradient quantization instead of one global quantization parameter (ablation) | INT8 training with ResNet-18 · ImageNet | Accuracy (%): 0.30% | Helped |
| Use magnitude-aware clipping to set gradient quantization parameters (ablation) | INT8 training with ResNet-18 · ImageNet | Accuracy (%): 0.33% | Helped |
| Combine vectorized quantization with magnitude-aware clipping (ablation) | INT8 training with ResNet-18 · ImageNet | Accuracy (%): 0.53% | Helped |
| Vary MCS hyperparameters k and A across the tested pairs (ablation) | INT8 training with ResNet-18 · ImageNet | Accuracy (%): 69.94; 70.01; 69.90; 69.82; 69.80; 69.91 | No clear effect |
| Use distribution-adaptive INT8 gradient quantization instead of UI8 | Faster-RCNN with ResNet-50 · COCO | mAP: 36.9 to 35.1 (UI8); 37.5 to 37.4 (Ours) | Helped |
| Use distribution-adaptive INT8 gradient quantization instead of UI8 | RetinaNet with ResNet-50 · COCO | mAP: 36.2 to 35.0 (UI8); 36.4 to 36.3 (Ours) | Helped |
| Quantize gradients with distribution-adaptive INT8 | ResNet3D-18 · Kinetics-400 | Top-1 accuracy (%): 51.0 -> 51.4 | Helped |
| Quantize gradients with distribution-adaptive INT8 | ResNet3D-50 · UCF-101 split-1 | Top-1 accuracy (%): 57.6 -> 57.5 | Hurt |
Closest papers by meaning
- Towards Unified INT8 Training for Convolutional Neural Network (CVPR 2020 · similarity 0.92)
- A Statistical Framework for Low-bitwidth Training of Deep Neural Networks (NeurIPS 2020 · similarity 0.88)
- Fixed-Point Back-Propagation Training (CVPR 2020 · similarity 0.87)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.86)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.86)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 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.