Optical Convolution Processing Based on an Amplified Fiber-Optic Recirculating Loop

Convolution processing is a key function in convolutional neural networks (CNNs). To increase the computational speed of a CNN, optical convolution processing (OCP) can be employed to leverage the high speed and parallelism of light. However, the scalability of most OCP schemes is poor. For a given system, the kernel size is fixed, or the system must be expanded with more components to handle convolution processing with larger kernel sizes. Here, we propose an approach to implementing OCP capable of handling various kernel sizes based on an amplified fiber-optic recirculating loop without changing the configuration of the system. For an input data sequence, the multiplication of the input data with a kernel having $n$ weights is implemented by applying the input data and the kernel weights to a MachZehnder modulator (MZM), to allow the weighted input data to recirculate in the amplified fiber-optic loop n-1 times. The weighted and time-delayed data are summed at a photodetector (PD), and the convolution operation is completed. A proof-ofconcept experiment is performed, in which the feature maps of handwritten digits from the MNIST dataset using different kernels at a speed of 16 giga operations per second (GOPS) are obtained.