Programmable Nanophotonics for Deep Learning and Neuromorphic Computing

用于深度学习和神经形态计算的可编程纳米光子学

• 批准号: RGPIN-2018-05249
• 负责人: Shastri, Bhavin
• 金额: $ 2.4万
• 依托单位: Queen's University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2019
• 资助国家: 加拿大
• 起止时间: 2019-01-01 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=684947
• 关键词: Programmable; Nanophotonics; Deep; Learning; Neuromorphic

SYNOPSIS. The birth of computers shaped 20th century society and science. After decades of exponential improvement, the performance of von Neumann architectures in speed, efficiency, and generality, has begun to run into fundamental limits, as the shrinking of transistors reaches its physical limits. As a result, the gap between current computing capabilities and computing needs is widening. This insufficiency is apparent in problems involving complex systems, big data, or real-time requirements. Forays into unconventional (non-von Neumann) computing have only been partially successful due to the limitations in bandwidth and energy consumption posed by metal interconnects.******VISION orm state-of-the-art microelectronic processors in energy efficiency and computational speeds by seven- and four orders-of-magnitude, respectively. Scientific objectives include: 1) devices thrustenergy efficient (attoJoule/operation) photonic neurons with graphene-based electro-optic modulators; 2) architectures thrustscalable and programmable silicon photonic neural network architectures; and 3) applications thrustphotonic processors for generalized neuromorphic computing tasks including deep learning and nonlinear optimization for model predictive control. This program focuses on these computing tasks as they are notoriously difficult to solve.******IMPACT. An experimentally-driven investigation of neuromorphic nanophotonics will serve as the first feasibility proof of using integrated photonics for scalable information processing. The proposed program has the potential to shape the emerging field of generalized compute engines beyond von-Neumann architectures and help redefine their physical limitations. The resulting technology has the potential to transform social, scientific, and technological sectors including self-navigating vehicles, bio-informatics, security, and big data. This research will contribute to the culture of innovation excellence across the scientific community and Canada. The multi-disciplinary nature of the program promises to foster collaborative ties across Canadian academic institutions and the private sector. The program's educational impact rests on uniquely positioning students for a readied workforce in academia or industry to drive tomorrow's advancements in photonics, applied physics, and engineering for 21st century challenges.

概要。计算机的诞生塑造了 20 世纪的社会和科学。经过几十年的指数级改进，随着晶体管的缩小达到其物理极限，冯·诺依曼架构在速度、效率和通用性方面的性能已经开始遇到基本限制。因此，当前计算能力与计算需求之间的差距正在扩大。这种不足在涉及复杂系统、大数据或实时要求的问题中尤为明显。由于金属互连带来的带宽和能耗限制，非常规（非冯诺依曼）计算的尝试仅取得了部分成功。******VISION 使最先进的微电子处理器在能源效率和计算速度方面分别提高了七个和四个数量级。科学目标包括：1）具有基于石墨烯的电光调制器的装置推力高效（attoJoule/操作）光子神经元； 2) 推力可扩展和可编程硅光子神经网络架构； 3）用于广义神经拟态计算任务的应用推力光子处理器，包括模型预测控制的深度学习和非线性优化。该程序专注于这些计算任务，因为它们非常难以解决。******影响。对神经形态纳米光子学的实验驱动研究将成为使用集成光子学进行可扩展信息处理的第一个可行性证明。拟议的计划有可能塑造冯诺依曼架构之外的通用计算引擎的新兴领域，并帮助重新定义其物理限制。由此产生的技术有可能改变社会、科学和技术领域，包括自动导航车辆、生物信息学、安全和大数据。这项研究将为整个科学界和加拿大的卓越创新文化做出贡献。该计划的多学科性质有望促进加拿大学术机构和私营部门之间的合作关系。该项目的教育影响力在于为学生提供独特的定位，为学术界或工业界做好准备的劳动力队伍，推动光子学、应用物理和工程领域的未来进步，应对 21 世纪的挑战。