祝贺易思成博士的全局光功率调配的光神经网络工作被Optics Letters期刊收录

中心博士易思成的工作——Enhancement of calculation accuracy of the integrated photonic tensor flow processer by global optical power allocation(利用全局光功率分配的方案提高集成光子张量流处理器的计算精度)的相关成果近期被Optics Letters期刊接收发表,该工作得到了国家自然科学基金(T2225023, 62205203)、国家重点研发计划(2019YFB2203700)的部分资助。目前光神经网络已成为一种有前途的硬件加速器,与数字电子驱动的机器学习相比,未来其具有显在的速度和能量优势。目前的光神经网络主要分为3种:相干光架构、衍射光神经网络和基于WDM权值广播的光神经网络。其中基于WDM权值广播的片上光卷积神经网络大多使用固定分光结构。为满足整体权重值计算,该结构造成很大一部分光功率未用于光学计算,降低光学计算的优势。本文提出了一种可以根据权重值权值大小进行全局光功率可调配的光学卷积神经网络架构,该架构能有效提升基于WDM权值广播的片上光卷积神经网络计算能力,在考虑片上延时线及分光损耗后,全局光功率可调配较固定分光架构在实验中测得计算精度提升1位以上,有利于未来片上实现大规模光卷积神经网络。

摘要: We present a global optical power allocation architecture, which can enhance the calculation accuracy of the integrated photonic tensor flow processer (PTFP). By adjusting the optical power splitting ratio according to the weight value and loss of each calculating unit, this architecture can efficiently utilize optical power so that signal-to-noise ratio of the PTFP is enhanced. In the case of considering the on-chip optical delay line and spectral loss, the calculation accuracy measured in the experiment is enhanced by more than 1 bit compared with the fixed optical power allocation architecture.