Research map / Architectures and efficiency / Quantization
Towards Unified INT8 Training for Convolutional Neural Network
Check 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 |
|---|---|---|---|
| Add direction-sensitive gradient clipping (ablation) | INT8 training with MobileNetV2 · CIFAR-10 | training stability / accuracy: NaN -> 93.02 | Helped |
| Use exponential learning-rate scaling instead of linear or quadratic scaling (ablation) | INT8 training with MobileNetV2 · CIFAR-10 | training stability | Helped |
| Vary scaling hyperparameters (α, β) among tested settings (ablation) | INT8 training with MobileNetV2 · CIFAR-10 | accuracy (%): 92.82; 93.28; 93.38; 93.27 | No clear effect |
| Update clipping less frequently (period 10, 100, or 1000 rather than 1) (ablation) | INT8 training with MobileNetV2 · CIFAR-10 | accuracy (%): 93.66 -> 93.07; 93.38; 92.75 | Hurt |
| Use INT8 training rather than full precision | ResNet-20 · CIFAR-10 | accuracy (%): 92.32 -> 91.95 | Hurt |
| Use INT8 training rather than full precision | MobileNetV2 · CIFAR-10 | accuracy (%): 94.39 -> 93.38 | Hurt |
| Use INT8 training rather than full precision | InceptionV3 · CIFAR-10 | accuracy (%): 94.89 -> 95.00 | Helped |
| Use INT8 training rather than FP8 training | ResNet-18 · ImageNet | accuracy (%): 67.34 -> 69.67 | Helped |
| Use INT8 training rather than WAGE | AlexNet with 2-bit weights · ImageNet | accuracy (%): 48.40 -> 51.28 | Helped |
| Use INT8 training rather than full precision | ResNet-50 · ImageNet | accuracy (%): 76.60 -> 76.34 | Hurt |
| Use INT8 training rather than full precision | Faster R-CNN / ResNet-50 · PASCAL VOC | mAP (%): 82.0 -> 81.9 | Hurt |
| Use INT8 training rather than full precision | RetinaNet / ResNet-50 · COCO | mAP (%): 36.9 -> 35.1 | Hurt |
Closest papers by meaning
- Distribution Adaptive INT8 Quantization for Training CNNs (AAAI 2021 · similarity 0.92)
- Scalable methods for 8-bit training of neural networks (NeurIPS 2018 · similarity 0.90)
- Post training 4-bit quantization of convolutional networks for rapid-deployment (NeurIPS 2019 · similarity 0.89)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.88)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.87)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · 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.