Supercontinuum-based photonic neural network for all-optical data classification

光子平台的全光处理在能量效率和传输带宽方面优于当前的电子硬件。因此，它为即将到来的社会经济进步（如物联网、大数据网络以及金融市场的新型全球服务）所带来的速度和容量要求提供了解决方案。所要求的设备是一个高度可重构的超连续统生成系统，它是实现一种新型的、通用的机器学习协处理器的不可或缺的工具包，用于以光速进行全光信息分类和路由。它将满足以5G/6G和bbb100 Gbit/s革命为代表的当前和未来电信基础设施的需求。具体来说，该系统构成了前馈型神经网络的光子硬件实现，能够处理存储在宽带脉冲的频谱相位和幅度中的信息。这是通过创建和互连光谱-时间模式（即虚拟网络节点），通过涉及超连续统生成的复杂非线性光学过程来实现的。因此，高光谱输出空间很容易分离，用于广泛的信息分类任务。我们的系统旨在在以下方面超越其他光子神经形态平台：1)处理速度，因为它可以不断适应传入的数据速率；2)功耗，与海外电缆链路中使用的标准电信信号中继器相当；3)学习能力。特别是，后者允许我们的神经网络处理器在训练系统处理多个高度多样化的任务时具有异常多功能性。这些包括a)电信数据和头部识别，b)激光显微镜图像数据的单镜头在线特征识别，以及c)复杂电信网络的在线色散监测。在我们的建议中，我们将处理拟议网络的实施和培训，以及其基准测试。所要求的设备包括一套成本效益高的现成光纤组件（转化为设备的可重复性和实用性），即：1)用于信息编码的宽带可编程相位和幅度滤波器，以及2)用于超连续谱生成的色散补偿和高度非线性光纤模块（都是维持偏振的），以及3)用于实现和基准全光相位编码的可重构激光系统。该捆绑包对于扩展当前的光学处理能力至关重要，PI的设备代表了最先进的技术，使用现成的电信设备的成本和功率效率高的设备。

All-optical processing in photonic platforms outperforms current electronic hardware in energy efficiency and transfer bandwidth. As such, it offers the solution to the speed and capacity requirements imposed by upcoming socio-economical advances such as the Internet of Things, Big Data networks, as well as novel global services for financial markets. The requested equipment a highly-reconfigurable supercontinuum generation system represents an indispensable toolkit to realize a novel, versatile machine learning co-processor for all-optical information classification and routing at the speed of light. It will serve the demands of today's and tomorrow's telecommunications infrastructure, represented by the 5G/6G and > 100 Gbit/s revolutions. Specifically, the system constitutes a photonic hardware implementation of a feed-forward-type neural network, enabling the processing of information stored in the spectral phases and amplitudes of broadband pulses. This is achieved by creating and interconnecting spectro-temporal modes (i.e., virtual network nodes) via the complex nonlinear-optical processes involved in supercontinuum generation. The resulting, hyperspectral output space is easily separable for a broad range of information-classification tasks. Our system aims to outperform other photonic neuromorphic platforms in 1) processing speeds, as it continuously adapts to the incoming data rate, 2) power consumption, comparable to that of standard telecom signal repeaters used in oversea cable links, and 3) learning capacity. In particular, the latter allows our neural network processor to be exceptionally versatile in training the system towards multiple, highly diverse tasks. These include a) telecom data and header recognition, b) single-shot in-line feature recognition of laser-microscope image data, and c) online dispersion monitoring of complex telecom networks. In our proposal, we will tackle the implementation and training of the proposed network, as well as its benchmarking. The requested equipment includes a cost-effective bundle of off-the-shelf fiber components (translating to device reproducibility and practicality), namely: 1) a broadband programmable phase and amplitude filter for information encoding, together with 2) dispersion compensating and highly nonlinear fiber modules (both polarization-maintaining) for supercontinuum generation, as well as 3) a reconfigurable laser system to implement and benchmark all-optical phase encoding for the tackled applications. The bundle is essential to extend current optical processing capabilities, where the PI's facilities represent the state-of-the-art, to cost- and power-efficient devices using off-the-shelf telecom equipment.