Abstract

Nowadays, photonic analog computing in the form of neuromorphic photonics or photonic neural networks has resulted in powerful hardware platforms that accelerate difficult tasks with marginal power consumption. Although linear transformations are efficiently carried out in the optical domain, nonlinear transformations are also of paramount importance. Here, we propose four wave mixing (FWM) process as an elegant solution for the generation of diverse nonlinear higher order products of an original signal. On one hand, these products can act as activation functions in different types of neural networks, either recurrent neural networks (RNNs) and reservoir computers or in any other type of brain-inspired frameworks. On the other hand, the same products can form the basis for polynomial regression which is a very powerful technique used in machine learning in different tasks i.e. for the prediction of COVID-19 transmission [1].

© 2023 IEEE