Brain-inspired photonic computing for efficient next-generation telecommunications networks

在物联网、大数据网络和智能应用的背景下，即将到来的社会经济进步正在给当前的信息和通信技术基础设施带来压力。这些需求正在达到传统技术解决方案的极限，这意味着需要新的概念来大规模处理数据。递归神经网络（RNN）是一种受大脑启发的机器学习范式，特别适合并行计算和模式识别，提供了几个数量级的处理加速。光子RNN承诺提供一个多功能平台，用于光速数据处理，其精度可与经典软件方法相媲美，并且占地面积和功耗更小。然而，由于需要电光转换，当前的实现方式通常在计算能力、系统简单性或处理速度方面缺乏。在拟议的项目中，我们将与我们的工业合作伙伴华为加拿大公司一起专注于三个主要目标，以克服这些限制：（i）首次展示功能性光子神经形态平台，利用多个自由度以前所未有的速度进行信息处理，（ii）其在电信数据处理方面的实际应用，以及（iii）开发高效紧凑的原型。我们的目标是利用片上非线性元件的非线性信道均衡和信号再生的任务。设想的原型将首次揭示适合大规模生产的光子RNN的全部潜力。该项目的成果将通过在集成光子学和机器学习的独特结合领域培训高素质人才（HQP），使加拿大市场受益，并有可能使HQP成为行业和学术界的未来领导者。我们技术的商业化将加强加拿大在人机交互和电信等高科技领域的作用，从而为科普即将到来的6 G和400 Gb/s系统等高密度传输标准的需求铺平道路。

Upcoming socio-economical advances in the context of the Internet of Things, big data networks, and smart applications are placing stress on current information and communication technology infrastructure. Such demands are reaching the limit of conventional technological solutions, implying a need for novel concepts in order to process data at scale. Recurrent neural networks (RNNs) are a brain-inspired machine learning paradigm that is especially suited for parallel computing and pattern recognition, offering a processing acceleration of several orders of magnitudes. Photonic RNNs promise a versatile platform for speed of light data processing comparable in accuracy with classical software approaches, at a reduced footprint and power consumption. However, current implementations are often lacking in terms of computational power, system simplicity, or processing speed, due to the need for electro-optical conversion. In the proposed project, we will, together with our industrial partner Huawei Canada, focus on three main objectives to overcome such limitations: (i) the first demonstration of a functional photonic neuromorphic platform exploiting multiple degrees of freedom for information processing at unprecedented speeds, (ii) its real-world application towards telecommunications data processing, and (iii) the development of an efficient and compact prototype. We aim to utilize on-chip nonlinear components to target the tasks of nonlinear channel equalization and signal regeneration. The envisioned prototype will, for the first time, reveal the full potential of photonic RNNs suitable for large-scale production. The project outcome will benefit the Canadian market through the training of highly-qualified personnel (HQP) in the uniquely combined fields of integrated photonics and machine learning, with the potential for the HQP to become future leaders in both industry and academia. The commercialization of our technology will strengthen the role of Canada in the high-tech sectors of human-machine interaction and telecommunications, thus paving the way to cope with the demands of upcoming high-density transmission standards such as 6G and 400 Gb/s systems.