Research map / Architectures and efficiency / Quantization
FATNN: Fast and Accurate Ternary Neural Networks
Peng Chen, Bohan Zhuang, Chunhua Shen
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 |
|---|---|---|---|
| add non-uniform step size quantization (ablation) | ternary LSQ baseline (ResNet-18) · ImageNet val | Top-1 accuracy: 64.7 -> 65.4 | Improved |
| add distribution calibration of the residual branch (ablation) | FATNN ternary (ResNet-18) · ImageNet val | Top-1 accuracy: 65.4 -> 66.0 | Improved |
| add distribution calibration of the residual branch (ablation) | full-precision ResNet-18 · ImageNet val | Top-1 accuracy: 69.8 -> 69.6 | Hurt |
| replace LSQ with FATNN quantization (ternary) | ResNet-18 · ImageNet val | Top-1 accuracy: 64.7 -> 66.0 | Improved |
| replace LSQ with FATNN quantization (ternary) | ResNet-34 · ImageNet val | Top-1 accuracy: 69.0 -> 69.8 | Improved |
| replace LSQ with FATNN quantization (ternary) | ResNet-50 · ImageNet val | Top-1 accuracy: 71.2 -> 72.4 | Improved |
| use FATNN ternary instead of LQ-Net 2-bit activation/binary weight | ResNet-18 · ImageNet val | Top-1 accuracy: 62.6 -> 66.0 | Improved |
| use FATNN ternary implementation instead of 2-bit implementation | ResNet-18 on Qualcomm 835 · execution time | speedup: 2.1 | Improved |
| use FATNN ternary implementation instead of 2-bit implementation | ResNet-34 on Qualcomm 821 · execution time | speedup: 2.0 | Improved |
| include extra bit operations needed by FATNN | ternary conv layers on Qualcomm 821 · layer-wise execution time | relative overhead: 0.2% to 10.3% | No clear effect |
| include extra bit operations needed by FATNN | ternary conv layers on Qualcomm 835 · layer-wise execution time | relative overhead: 0.2% to 32.0% | Mixed |
| include extra bit operations needed by FATNN | ternary conv layers on Nvidia 2080Ti · layer-wise execution time | relative overhead: 1.7% to 10.3% | No clear effect |
Closest papers by meaning
- Trained Ternary Quantization (ICLR 2017 · similarity 0.91)
- TRQ: Ternary Neural Networks With Residual Quantization (AAAI 2021 · similarity 0.90)
- HitNet: Hybrid Ternary Recurrent Neural Network (NeurIPS 2018 · similarity 0.89)
- RTN: Reparameterized Ternary Network (AAAI 2020 · similarity 0.89)
- HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS (ICLR 2025 · similarity 0.87)
- An Efficient Matrix Multiplication Algorithm for Accelerating Inference in Binary and Ternary Neural Networks (ICML 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.
