Research map / Architectures and efficiency / Quantization
Is Integer Arithmetic Enough for Deep Learning Training?
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 |
|---|---|---|---|
| reduce bit-width from int8 to int7 (ablation) | ResNet18 · CIFAR10 | accuracy: 94.8 -> 94.7 | No clear effect |
| reduce bit-width from int8 to int6 (ablation) | ResNet18 · CIFAR10 | accuracy: 94.8 -> 94.47 | No clear effect |
| reduce bit-width from int8 to int5 (ablation) | ResNet18 · CIFAR10 | accuracy: 94.8 -> 88.5 | Hurt |
| reduce bit-width to int4 (ablation) | ResNet18 · CIFAR10 | accuracy: 94.8 -> Diverges | Hurt |
| replace float training with int8 integer training | ResNet18 · ImageNet | top-1 accuracy: 69.75 -> 69.25 | No clear effect |
| replace float training with int8 integer training | MobileNetV2 · ImageNet | top-1 accuracy: 71.87 -> 72.80 | No clear effect |
| replace float training with int8 integer training | ResNet18 · CIFAR100 | top-1 accuracy: 76.14 -> 75.38 | No clear effect |
| fine-tune ViT-B with int8 integer pipeline | ViT-B-16-224 · CIFAR10 | top-1 accuracy: 99.10 -> 98.80 | No clear effect |
| int8 training of DeepLab-V1 | DeepLab-V1 · PASCAL VOC 2012 | mIOU: 75.00 -> 74.73 | No clear effect |
| int8 training of DeepLab-V2 | DeepLab-V2 · PASCAL VOC 2012 | mIOU: 77.71 -> 77.65 | No clear effect |
| int8 training of DeepLab-V1 | DeepLab-V1 · MS COCO 10K | mIOU: 34.70 -> 34.80 | No clear effect |
| int8 training of Faster R-CNN | Faster R-CNN · dataset 1 (COCO 10K) | mAP: 37.80 -> 37.40 | No clear effect |
Closest papers by meaning
- Training Deep Neural Networks with 8-bit Floating Point Numbers (NeurIPS 2018 · similarity 0.88)
- Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference (CVPR 2018 · similarity 0.88)
- Winning Both the Accuracy of Floating Point Activation and the Simplicity of Integer Arithmetic (ICLR 2023 · similarity 0.88)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.84)
- A Fast Convoluted Story: Scaling Probabilistic Inference for Integer Arithmetics (NeurIPS 2024 · similarity 0.84)
- Scaling Laws for Floating–Point Quantization Training (ICML 2025 · similarity 0.83)
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.