Research map / Architectures and efficiency / Quantization
F8Net: Fixed-Point 8-bit Only Multiplication for Network Quantization
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 F8Net instead of the strongest listed quantization baseline | ResNet18 · ImageNet | Top-1 accuracy: 71.1 -> 71.1 | No clear effect |
| Use F8Net instead of SAT | MobileNet V1 · ImageNet | Top-1 accuracy: 72.6 -> 72.8 | Helped |
| Use F8Net instead of SAT | MobileNet V2b · ImageNet | Top-1 accuracy: 72.5 -> 72.6 | Helped |
| Use F8Net instead of OMPQ | ResNet18 tiny fine-tuning · ImageNet | Top-1 accuracy: 72.3 -> 72.4 | Helped |
| Use F8Net instead of HAWQ-V3 | ResNet50b, Baseline #1 tiny fine-tuning · ImageNet | Top-1 accuracy: 77.5 -> 77.6 | No clear effect |
| Use F8Net instead of HAWQ-V3 | ResNet50b, Baseline #2 tiny fine-tuning · ImageNet | Top-1 accuracy: 78.1 -> 78.1 | No clear effect |
| Restrict the fractional-length search space to {6,7,8} instead of 0–8 (ablation) | F8Net on ResNet50 · ImageNet | Top-1 accuracy: 77.6 -> 72.4 | Hurt |
| Allow batch normalization to update during tiny fine-tuning instead of keeping it fixed (ablation) | F8Net tiny fine-tuning · ImageNet | Top-1 accuracy | No clear effect |
| Use private fractional lengths for sibling layers instead of sharing fractional length (ablation) | F8Net residual blocks · ImageNet | Top-1 accuracy | Helped |
| Apply weight decay to depthwise convolution and BN layers (ablation) | MobileNet V1 quantization · ImageNet | validation error | Helped |
| Apply weight decay to depthwise convolution or BN weights (ablation) | MobileNet V2 quantization · ImageNet | validation error | Hurt |
Closest papers by meaning
- FP8 Quantization: The Power of the Exponent (NeurIPS 2022 · similarity 0.91)
- Scalable methods for 8-bit training of neural networks (NeurIPS 2018 · similarity 0.90)
- Towards Unified INT8 Training for Convolutional Neural Network (CVPR 2020 · similarity 0.88)
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.88)
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.88)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.88)
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.