Related Experiment Video
Updated: Jul 31, 2026
Real-time Iontophoresis with Tetramethylammonium to Quantify Volume Fraction and Tortuosity of Brain Extracellular Space
Published on: July 24, 2017
CHARLES: A C++ fixed-point library for Photonic-Aware Neural Networks
Emilio Paolini1, Lorenzo De Marinis2, Luca Maggiani3
1Scuola Superiore Sant'Anna, Pisa, 56124, Italy; National Research Council of Italy - Institute of Electronics, Information Engineering and Telecommunications (CNR-IEIIT), Pisa, 56122, Italy; Sma-RTy Italia Srl, Carugate, 20061, Italy.
We introduce CHARLES, a C++ library for simulating Photonic-Aware Neural Networks (PANNs). Fixed-point training significantly improves PANN inference accuracy on low bitwidths compared to floating-point training.
Area of Science:
- Computer Science
- Electrical Engineering
- Photonics
Background:
- Photonic-Aware Neural Networks (PANNs) must account for hardware constraints, particularly low numerical precision.
- Simulating PANNs requires tools that handle low-precision computations.
Purpose of the Study:
- To present CHARLES, a C++ library for simulating PANNs.
- To compare the impact of fixed-point versus floating-point training on PANN inference accuracy under quantization.
Main Methods:
- Developed CHARLES, a C++ library supporting fixed-point inference and both floating-point/fixed-point training.
- Evaluated CHARLES using Iris, MNIST, and Fashion-MNIST datasets.
- Compared accuracy loss from quantization after floating-point training versus fixed-point training.
Main Results:
- Fixed-point training outperforms floating-point training for PANN inference on low bitwidths.
- Floating-point training followed by quantization leads to significant accuracy loss.
- Fixed-point training minimizes accuracy loss, achieving 90.4% (MNIST) and 68.1% (Fashion-MNIST) at 6 bits.
Conclusions:
- Fixed-point training is crucial for effective PANN deployment on hardware with limited precision.
- CHARLES provides a validated tool for PANN simulation and numerical format optimization.
