Research map / Architectures and efficiency / CNN design
Mixed Precision Training of Convolutional Neural Networks using Integer Operations
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 mixed-precision DFP16 instead of FP32 training | ResNet-50 · ImageNet-1K | Top-1 accuracy: 75.70\% -> 75.77\% | Helped |
| Use mixed-precision DFP16 instead of FP32 training | ResNet-50 · ImageNet-1K | Top-5 accuracy: 92.78\% -> 92.84\% | Helped |
| Use mixed-precision DFP16 instead of FP32 training | GoogLeNet-v1 · ImageNet-1K | Top-1 accuracy: 69.26\% -> 69.34\% | Helped |
| Use mixed-precision DFP16 instead of FP32 training | GoogLeNet-v1 · ImageNet-1K | Top-5 accuracy: 89.31\% -> 89.31\% | No clear effect |
| Use mixed-precision DFP16 instead of FP32 training | VGG-16 · ImageNet-1K | Top-1 accuracy: 68.23\% -> 68.12\% | Hurt |
| Use mixed-precision DFP16 instead of FP32 training | VGG-16 · ImageNet-1K | Top-5 accuracy: 88.47\% -> 88.18\% | Hurt |
| Use mixed-precision DFP16 instead of FP32 training | AlexNet · ImageNet-1K | Top-1 accuracy: 57.43\% -> 56.94\% | Hurt |
| Use mixed-precision DFP16 instead of FP32 training | AlexNet · ImageNet-1K | Top-5 accuracy: 80.65\% -> 80.06\% | Hurt |
| Use mixed-precision DFP16 instead of FP32 | ResNet-50 · ResNet-50 training | training throughput: 1.8X | Helped |
| Use mixed-precision DFP16 versus a tuned FP32 version | ResNet-50 · ResNet-50 training | speedup: 1.6X | Helped |
Closest papers by meaning
- Mixed Precision Training (ICLR 2018 · similarity 0.88)
- Training Deep Neural Networks with 8-bit Floating Point Numbers (NeurIPS 2018 · similarity 0.87)
- DQT: Dynamic Quantization Training via Dequantization-Free Nested Integer Arithmetic (AAAI 2026 · similarity 0.86)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.86)
- Towards Unified INT8 Training for Convolutional Neural Network (CVPR 2020 · similarity 0.86)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · similarity 0.85)
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.