Optical Neural Network With Complementary Decomposition to Overcome the Phase Insensitive Constrains

Abstract

Photonic integrated circuits have potentials to advance artificial neural networks by providing ultra-fast and high power efficient computation resources. In many of optical neural network (ONN) architectures, data are encoded in light amplitudes, using magnitude and phase to denote the absolute value and the sign. However, the phase insensitivity of optical components such as saturable absorbers, semiconductor optical amplifiers and photodetectors can lead to limitations in ONN designs, including nonlinear activation and inference results readout. In this work, we propose a complementary decomposition approach to overcome the phase insensitive problem. Based on that, we further present a optical ReLU function implementation. The proposed designs can scale to multi-layer neural networks by cascading linear and nonlinear units. These optical neural networks do not require repeated bidirectional digital-analog conversions and optical-electronic conversions between each layers as well as reference signals for homodyne detection. These features could provide a potential implementation of compact, scalable, high efficiency ONNs. We evaluate the concept with a machine learning task. The numerical results demonstrate that the proposed design can tackle the aforementioned phase insensitive problem.