PIC: Hybrid Silicon Electronic-Photonic Integrated Neuromorphic Networks

神经形态计算是人工智能的一个子领域，它实现了受大脑学习过程启发的物理架构。使用电子集成电路技术来实现神经网络结构已经做出了重大努力。然而，纯电子硬件不适合对现代信息世界至关重要的高带宽应用。相比之下，由于光的高带宽、高速度和低能耗，互联网由光子技术(激光、电光调制器和光电探测器)供电。因此，本项目的目标是实现利用光的高性能神经网络。这些光子神经网络将被集成在一个光子芯片上，以实现可扩展和高效的体系结构。然而，为了建立超越当今最先进水平的神经网络，由于围绕光子存储和放大的挑战，还有必要利用电子技术，这两者都是实现通用神经网络的关键。这种将电子学和光子学结合在一起的混合方法，使研究最广泛的问题成为可能。除了这些研究目标外，该项目还是一项跨学科活动，将为未来的科学和工程专业人员提供技术培训。将有外展活动，将研究带给K-12、本科生和研究生。来自代表性不足背景的学生将通过提供实验室参观和动手活动来积极参与。最后，将利用AIM光子学院的教育举措来传播该项目中的研究成果。总的来说，这个项目将通过在自主系统、视觉系统、信息网络、网络安全、机器人和其他高带宽应用中的应用来影响更广泛的社区。本项目旨在解决两个基本问题，i)光子学如何最大限度地发挥计算领域的功能？ii)什么神经形态算法可以利用光子学解决广泛的问题？本项目的总体目标是展示混合硅电子-光子集成神经形态网络。该模型利用光干涉的能力实现高性能的神经形态计算网络。神经网络的光子实现提供了固有的优势，即光可以轻松地执行传统上在纯电子实现中难以完成的计算任务(例如，傅立叶变换可以通过光学方式完成，只需让光通过透镜)。本文提出的底层集成光电子网络使用多模干涉耦合器作为神经核心(Neuro-MMI)，以便在紧凑的空间内实现多个输入和输出之间的干扰。主要研究人员提出了实现围绕MMI核心的神经网络中权重的可重构性。Neuro-MMI内核将与光电非线性阈值电路(连同电子存储器)集成，以实现不同类型的神经网络(前馈神经网络和递归神经网络)。将研究有源神经MMI以实现片上学习，并将研究这些拓扑所固有的新的学习规则。此外，还将探索使用波分复用来实现密集连接和并行，以最大限度地提高网络性能。拟议的光子-电子混合网络的一个独特功能是可以在单个芯片上在前馈和递归神经网络之间进行重新配置。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Neuromorphic computing is a sub-field of artificial intelligence that implements physical architectures inspired by the learning processes in the brain. There have been significant efforts to realize neural network architectures using electronic integrated circuit technology. However, electronic-only hardware is not suitable for high bandwidth applications critical to a modern information world. In contrast, the internet is powered by photonic technologies (lasers, electro-optic modulator and photodetectors) because of light's high bandwidth, speed and low energy consumption. Consequently, this project aims to realize high performance neural networks that utilize light. These photonic neural networks will be integrated on a photonic chip in order to realize scalable and efficient architectures. However, in order to build neural networks that transcend today's state-of-art, it is necessary to also leverage electronics due to the challenges surrounding photonic memory and amplification, both of which are key to realizing a general purpose neural network. This hybrid approach, where electronics and photonics would be integrated together, enables the investigation of the broadest class of problems. In addition to these research aims, this project is an interdisciplinary activity that will provide technical training for future science and engineering professionals. There will be outreach activities that bring the research to K-12, undergraduate, and graduate students. Students from underrepresented backgrounds will be actively engaged by providing lab visits with hands-on activities. Lastly, the education initiatives of AIM Photonics Academy will be leveraged to disseminate the research in the project. Overall this project will impact the broader community with applications in autonomous systems, vision systems, information networks, cybersecurity, robotics and other high bandwidth applications.This project aims to address two fundamental questions, i) How can photonics maximize functionality in the compute domains?, ii) What neuromorphic algorithms can solve a broad class of problems using photonics?The overall goal of this project is to demonstrate hybrid silicon electronic-photonic integrated neuromorphic networks. The proposed paradigm leverages the power of optical interference to realize high performance neuromorphic computing networks. Photonic implementations of neural networks offer the inherent advantage that light can easily perform computational tasks that are traditionally challenging to do in electronic-only implementations (e.g. a Fourier transform can be done optically by simply passing light through a lens). The underlying integrated photonic-electronic network proposed here utilizes a Multimode interference coupler as a neural core (Neuro-MMI) in order to realize interference between multiple inputs and outputs in a compact footprint. The principal investigators propose to realize reconfigurability of the weights in the neural network wrapped around the MMI core. The Neuro-MMI core will be integrated with optoelectronic nonlinear thresholding circuits (along with electronic memory) to realize different classes of neural networks (feed forward neural networks and recurrent neural networks). Active Neuro-MMI's will be studied to realize on-chip learning and new learning rules will be investigated that are inherent for these topologies. Furthermore, the use of wavelength division multiplexing will be explored to achieve dense connectivity and parallelism in order to maximize the performance of the networks. One unique feature of the proposed hybrid photonic-electronic network is the reconfigurability to switch between feed forward and recurrent neural networks on a single chip.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.