Research map / Architectures and efficiency / Quantization
Fixed-Point Back-Propagation Training
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 |
|---|---|---|---|
| Use layer-wise adaptive fixed-point training instead of float32 | AlexNet · ImageNet | Top-1 accuracy: 58.0 -> 58.22 | Helped |
| Use layer-wise adaptive fixed-point training instead of float32 | VGG16 · ImageNet | Top-1 accuracy: 71.0 -> 70.6 | Hurt |
| Use layer-wise adaptive fixed-point training instead of float32 | ResNet50 · ImageNet | Top-1 accuracy: 76.4 -> 76.2 | Hurt |
| Use layer-wise adaptive fixed-point training instead of float32 | MobileNet-v2 · ImageNet | Top-1 accuracy: 71.8 -> 70.5 | Hurt |
| Use layer-wise adaptive fixed-point training instead of float32 | IMG Res101 · ImageNet Detection | mAP: 44.1 -> 44.4 | Helped |
| Use layer-wise adaptive fixed-point training instead of float32 | deeplab-v1 · VOC | meanIoU: 70.1 -> 69.9 | Hurt |
| Adapt backward-pass bit-width by layer instead of using fixed int16 (ablation) | Sockeye RNN · WMT’17 news translation | word-level accuracy: int16 gradually results in 2% loss; 61.97% -> 62.05% | Helped |
| Use adaptive fixed-point training instead of float32 | Transformer · WMT’16 Multi30k | word-level accuracy: 54.13% -> 55.54% | Helped |
| Use adaptive fixed-point training | AlexNet · ImageNet | training speedup: 2.52 times | Helped |
| Quantize the fc2 activation gradient to int8 rather than float32 (ablation) | AlexNet · ImageNet | Top-1 accuracy: 58.00% -> 48.27% | Hurt |
| Quantize the fc2 activation gradient to int12 rather than float32 (ablation) | AlexNet · ImageNet | Top-1 accuracy: 58.00% -> 50.30% | Hurt |
| Quantize the fc2 activation gradient to int16 rather than float32 (ablation) | AlexNet · ImageNet | Top-1 accuracy: 58.00% -> 58.28% | No clear effect |
Closest papers by meaning
- Adaptive Quantization of Neural Networks (ICLR 2018 · similarity 0.89)
- Deep Neural Network Quantization via Layer-Wise Optimization Using Limited Training Data (AAAI 2019 · similarity 0.89)
- Adaptive Quantization for Deep Neural Network (AAAI 2018 · similarity 0.89)
- Per-Tensor Fixed-Point Quantization of the Back-Propagation Algorithm (ICLR 2019 · similarity 0.89)
- Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats (ICLR 2023 · similarity 0.88)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 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.