Neuromorphic Network with Photonic Weighting and Photoelectronic Nonlinear Activation Based on SOA and APD

Abstract

Photonic neuromorphic computing is emerging as a promising approach for low-latency, energy-efficient nonvon Neumann computing systems. Nonlinear activation functions are key components of photonic neural networks. However, the current implementation of nonlinear activation functions faces issues such as high power consumption, high threshold, and limited bandwidth. Implementing efficient nonlinear activation functions in photonic neural networks remains a significant challenge. Here, we propose and demonstrate for the first time a photonic-electronic neural network consisting of semiconductor optical amplifiers (SOAs) as photonic synapses and InGaAs avalanche photodiodes (APDs) with trans-impedance amplifiers as photoelectronic nonlinear neurons. The APD achieves a tanh-like nonlinear activation function while performing photoelectronic conversion without additional electronic processing. It has a large operating bandwidth and a response frequency of 8 GHz, with a threshold input optical power of −30 dBm. Utilizing the proposed linear weighting mechanism of SOAs, low-power input signals can be weighted while being provided with gain. Embedding the measured nonlinear activation function into convolutional neural networks with two convolutional layers results in inference accuracies of 98.485% and 91.2% for the classification tasks of the MNIST handwritten digit image data set and the Fashion-MNIST image data set, respectively. Furthermore, leveraging the linear weighting mechanism of SOAs and nonlinear activation of APDs, we have successfully accomplished hardware inference tasks on the proposed network architecture. In addition, the computational speed and energy efficiency of the proposed network architecture are evaluated. We believe this work holds promise for advancing the development of monolithically integrated large-scale photonic-electronic intelligent neuromorphic systems in the future.