Research map / Architectures and efficiency / Quantization
Toward INT4 Fixed-Point Training via Exploring Quantization Error for Gradients
Dohyung Kim, Junghyup Lee, Jeimin Jeon, JAEHYEON MOON, Bumsub Ham
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| adaptively update gradient clipping factor to keep large-gradient error small, instead of fixed gamma=1.0 (ablation) | 4-bit FXP training ResNet-20 · CIFAR-100 test, 4/4/4 | top-1 accuracy: 61.1 -> 65.0 | Improved |
| adaptive interval update vs fixed gamma=1.0 on ResNet-18 (ablation) | FXP training ResNet-18 · ImageNet val, 4/4/4 | top-1 accuracy: 63.3 -> 67.1 | Improved |
| minimize quantization error for entire gradients (DSGC) instead of large gradients (ablation) | DSGC · CIFAR-100 test, 4/4/4 | top-1 accuracy: 61.1 (baseline) vs 24.3 (DSGC) | Hurt |
| use small fixed clipping factor gamma=0.6 on ImageNet ResNet-18 (4/4/4) (ablation) | baseline · ImageNet val, 4/4/4 | top-1 accuracy: 63.3 -> 50.2 | Hurt |
| use adaptive interval vs piecewise FXP format IQB | FXP training ResNet-18 · ImageNet val, 4/4/4 | top-1 accuracy: 59.9 -> 67.1 | Improved |
| apply interval update to object detection (ablation) | Faster R-CNN ResNet-50 · COCO test, 8/8/8 | mAP: 36.6 -> 37.0 | Improved |
| apply interval update to super-resolution EDSR (ablation) | EDSR 4/4/4 · Set5 x4 | PSNR: 31.82 -> 31.93 | Improved |
| use interval update algorithm vs plain baseline in runtime (ablation) | FXP training · CIFAR-100 latency | latency | No clear effect |
| adaptive update at 8/8/8 on ResNet-50 vs DSGC | FXP training ResNet-50 · ImageNet val, 8/8/8 | top-1 accuracy: 76.3 (DSGC) vs 76.0 (ours) | No clear effect |
Closest papers by meaning
- Network Quantization With Element-Wise Gradient Scaling (CVPR 2021 · similarity 0.91)
- Fixed-Point Back-Propagation Training (CVPR 2020 · similarity 0.90)
- Accurate Neural Training with 4-bit Matrix Multiplications at Standard Formats (ICLR 2023 · similarity 0.90)
- Data-Free Quantization Through Weight Equalization and Bias Correction (ICCV 2019 · similarity 0.90)
- Post-training Quantization with Multiple Points: Mixed Precision without Mixed Precision (AAAI 2021 · similarity 0.90)
- AMPA: Adaptive Mixed Precision Allocation for Low-Bit Integer Training (ICML 2024 · similarity 0.89)
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.
