The Hybrid Photonic Convolutional Neural Networks Based on SOA and FP-SA

Photonic convolutional neural networks (PCNNs) are gaining attention owing to their potential to overcome the von Neumann bottleneck. Here, we propose and demonstrate experimentally and numerically for the first time a hybrid PCNN consisting of semiconductor optical amplifiers (SOAs) as synapses and integrated quantum-well Fabry-Perot lasers with saturable absorbers (FP-SAs) as photonic neurons. In the experiment, both convolutional computations and nonlinear activation functions are performed in the optical domain. The time series is separated into discrete slots, each of which corresponds to a dot product operation within a sliding window between the convolution kernel matrix and the input matrix. The fast time-varying weights property of SOAs is used to dynamically configure the convolution kernels. The FP-SA laser neurons can be activated nonlinearly by input optical signals with multiple wavelengths. In addition, the performance metrics of the network architecture, such as computational efficiency, energy consumption, and throughput, are evaluated. Furthermore, the hybrid PCNN has been effectively utilised to compute image gradient magnitude for edge detection. Besides, we implement the handwritten digit classification task using the hybrid PCNN. The proposed network exhibits robustness in dealing with noisy images. The proposed approach c provide a promising scheme to construct a large-scale integrated hybrid PCNN chip.