Quantization-Aware Training for Mixed Precision Photonic Neural Networks

Manos Kirtas, Nikolaos Passalis, Athina Oikonomou, Miltiadis Moralis-Pegios, George Giamougiannis, Apostolos Tsakyridis, George Mourgias-Alexandris, Nikolaos Pleros, Anastasios Tefas

Abstract: Neuromorphic photonics is among the most promising approaches for Deep Learning (DL) hardware accelerators with recent state-of-the-art mixed precision configurations lowering even further the energy consumption while increasing simultaneously the computation rate. To this end, we propose a novel stochastic mixed precision quantization-aware training scheme oriented to Photonic Neural Networks (PNNs) that gradually lowers the bit resolution of layers based on their position and the observed tolerance to quantization noise. The effectiveness of the proposed method is demonstrated on an image classification task applying a ResNet8 architecture and a financial time series forecasting applying a Recurrent Neural Network (RNN) using two photonic configurations.