[ME1] Training for Photonic Neural Networks

11:45 AM - 12:00 PM JST(2:45 AM - 3:00 AM UTC)

[ME1-3] Mitigating Noise Effects in Photonic Neural Networks using Adaptive Quantization

○Emilio Paolini¹, Lorenzo De Marinis¹, Peter Seigo Kincaid¹, Luca Valcarenghi¹, Giampiero Contestabile¹, Giannis Roumpos^2,4, Miltiadis Moralis-Pegios^3,4, Nikos Pleros^3,4, Nicola Andriolli⁵ (1. Scuola Superiore Sant'Anna (Italy), 2. Dept. of Physics, Aristotle University of Thessaloniki, 541 24 Thessaloniki, Greece (Greece), 3. Dept. of Informatics, Aristotle University of Thessaloniki, 541 24 Thessaloniki, Greece (Greece), 4. Wireless and Photonic Systems and Networks Research Group, Center for Interdisciplinary Research and Innovation, Aristotle University of Thessaloniki, 541 24 Thessaloniki, Greece (Greece), 5. Department of Information Engineering, University of Pisa, 56122 Pisa, Italy (Italy))

Keywords:

Photonics for Computing and Deep Learning Applications

Photonic neural networks offer energy-efficiency but suffer from noise-induced low precision. We propose AQ-PANN, which learns a quantization step size to mitigate noise. Experiments on SVHN show strong performance across bitwidths under different noise levels.