Non-volatile and ultra-fast photonic vector accelerator with optical phase change materials and integrated microcomb

Abstract

Convolutional neural network (CNN) has attracted widespread attention in image feature extraction and speech recognition owing to greatly reducing the complexity of model parameters and the number of weights, but it cannot be separated from the support of hardware accelerator. The limitations of electronic devices in terms of power, speed, and size make it difficult for current electron accelerators to meet the computational power requirements of future large-scale convolution operations. Here, we proposed a photonic vector architecture. This structure combines time, space and wavelength, and the non-volatile phase change material and the integrated microcomb form an optical matrix multiplier to realize memory calculation, thus reducing the energy consumption of reading weight data. The tooth spacing of the integrated microcomb is more than 100 GHz, and the microcomb coverage is from 1510 nm to 1610 nm. Finally, we replace the weight values in the CNN with the optimal weight values that the optics can achieve. The final recognition accuracy reached 97.04%, which is comparable to the efficiency of the first electronic equipment. Our results could be helpful for the development of non-volatile and ultra-fast optical neural network (ONN) with feathers of low energy consumption and high integration.