Image classification with a simple photonic perceptron based on heterodyne detection

The use of photonic architectures to build neural networks is a fast-growing field, both because these systems offer huge computing potential, and because the power consumption of photonic systems is much lower than that of their electronic equivalents. In this context, we have developed a simple architecture based on a simple heterodyne interferometer, enabling vector multiplication. This architecture enables us to process images in the time domain, and perform image classification tasks. More importantly, the building block we propose can be generalized to multi-heterodyne architectures, enabling ultra-high rate matrix computation for neural network layers, and accelerators for convolutional neural networks.