基于INTERCONNECT的光子神经网络的设计

Design of Photonic Neural Network Based on Interconnect

近年来,代表人工智能(Artificial Intelligence,AI)的神经网络技术正朝着高速低功耗的方向发展。然而,由于电子器件的固有极限,传统电子神经网络功率效率与计算速度难以得到进一步提高。而光子神经网络能够把光电子技术与神经网络模型有机地结合,提供了突破这一瓶颈的有效手段。该文介绍了基于INTERCONNECT软件搭建的光子神经网络Gri Net,分析该网络的线性计算结构和非线性激活单元,并实现对手写数字的识别,准确率可以达到90%。不同于以往单纯的数学理论实现,该文利用物理器件搭建了光子神经网络,成果可用于光学神经网络的训练算法开发和光子芯片的高效识别任务,有较强的现实意义。

In recent years, the neural network technology representing Artificial Intelligence(AI) is developing towards high speed and low power consumption. However, due to the inherent limit of electronic devices, it is difficult to improve the power efficiency and computing speed of traditional electronic neural networks. Photonic neural network can combine photo electronic technology and neural network model organically, providing an effective means to break through this bottleneck. In this paper, the method and process of realizing GridNet based on optical device simulation in INTERCONNECT software are introduced, and its structure and mode characteristics are analyzed, so as to realize the recognition of handwritten numbers in the whole network, and the accuracy can reach 90%. Different from the previous simple mathematical theory, this paper uses physical devices to build the photonic neural network. The results wil be used for the development of the algorithm and training strategy of the anti-error optical neural network and the efficient recognition task of the photon chip,which has strong practical significance.