基于单片集成放大反馈半导体激光器的超快光子储备池计算

Reservoir computing (RC) is a brain-inspired new concept for information processing, and it can be used in many fields such as time series prediction, channel equalization. At present, the delay-based RC systems have attracted much attention due to easy implementation and relatively simple training, but it cannot satisfy the requirement for high-speed information processing since the processing speed of existingly delay-based RC systems is still at a level of M/s. Therefore, it is urgent to develop novel RC systems with ultra-fast (above G/s) processing speed. In this project, we plan to utilize monolithically integrated amplified feedback semiconductor lasers (MIAF-SLs) as reservoirs to realize RC with a capacity of G/s high-speed processing. Firstly, the dynamic characteristics of MIAF-SL taken as a reservoir in RC system is investigate deeply, and the focal points are the influences of system parameters on the consistency and memory capacity of the reservoir. Secondly, the structures of input and output layers matched with the reservoir are determined and constructed. Thirdly, for different ultra-fast information processing tasks such as time series prediction and channel equalization, the output weights for virtual nodes are calculated through training the RC system; Finally, the performance of RC system is evaluated, and the optimized RC system scheme is proposed for realizing high accuracy (error < 10 %) and ultra-fast information processing. Through the implementation of this project, it is expected to develop relevant novel technologies for ultra-fast RC, which is helpful to drive innovative development of artificial intelligence field in our country.

储备池计算(RC)是一种模仿大脑对信息进行处理的新型计算概念,可用于预测、信道均衡等领域,其中延时RC系统因易于实现且训练过程相对简单而备受关注.目前报道的延时RC系统的处理速度尚停留在M/s水平,达不到高速信息处理的要求,因此亟待研发超快(G/s以上)的RC系统.本项目拟采用单片集成放大反馈半导体激光器(MIAF-SLs)作为储备池以实现具有G/s高速处理能力的RC,具体为:对作为储备池的MIAF-SLs的动力学特性进行研究,着重关注系统参量对储备池的一致性、记忆能力等特性的影响;确定并搭建与该储备池匹配的RC系统输入层与输出层;针对不同的超快信息处理任务(时间序列预测、信道均衡等)对RC系统进行训练,得到各虚拟节点的输出权值;对RC系统的性能进行评估并优化,最终实现高准确率(误差<10%)的超快信息处理.本项目可望研发出超快RC相关新技术,服务于我国人工智能领域的创新发展.

储备池计算(RC)是一种模仿大脑对信息进行处理的新型计算概念，可用于预测、信道均衡等领域，其中延时RC系统因易于实现且训练过程相对简单而备受关注。为了实现对信息进行高速有效地处理，研发超快、高准确率的延时RC系统具有重要意义。我们在基于延时反馈半导体激光器作为储备池以构建高性能延时RC系统方面开展了相关研究探索。完成的主要研究内容及取得的重要结果有：（1）理论和实验研究了多种半导体激光器（DFB-SL、VCSEL、量子点激光器、带间级联激光器）在延时反馈作用下的非线性动力学特性，分析了系统参量对激光器非线性动力学特性的影响。（2）基于延时反馈半导体激光器构建延时RC系统，对该系统处理单任务的性能进行了理论研究，并搭建了相应的实验研究平台进行了实验研究，同时对进一步提升系统性能进行了探索。（3）基于延时反馈半导体激光器构建了能并行处理双任务的延时RC系统，并对系统性能进行了深入分析。（4）对基于半导体激光器的光子神经网络系统进行了初步的理论探索，对系统中神经元的spike编码、传输及其存储特性进行了数值分析。通过该项目的实施，在基于半导体激光器构建高性能延时RC系统方面取得了一些进展和突破，相关成果有望在人工智能领域得到应用。