Noise-Resistant Optical Implementation of Analogue Neural Networks

Abstract

Analogue artificial neural networks are widely considered as promising computational models that more closely imitate the information processing capabilities of the human brain compared to digital neural networks. The significant computation power and the much reduced power consumption per operation make the analogue implementation of neural networks very attractive. There is an active research on artificial neural networks (ANNs) implementation using both analogue photonic and electronic hardware [1] – [4] . However, compared to digital realisations the conventional analogue systems are more sensitive to the noise that is inevitably present in practical implementations [2] , [3] . Noise properties in ANNs have been studied both in the electronic and photonic domains. However, photonic ANNs are much less investigated compared to the electronic implementations, for which some training techniques have been proposed to enhance ANNs robustness against noise [1] , [4] .