Large-scale photonic-electronic integration for next generation neuromorphic computing systems

由于深度学习等新型机器学习算法的出现，神经形态计算在过去十年中获得了巨大的动力。人工神经网络处于这场革命的最前沿，其高效的硬件实现对许多科学和技术领域提出了重大挑战。神经网络计算所面临的主要问题是处理由密集连接的神经元层之间的并行信息流引起的大矩阵乘法。光子学为光学系统固有的并行操作等苛刻任务提供了独特的优势。近年来，人们提出了许多基于自由空间光学设置的方案，这些方案已经成功地实现了具有数十万个神经元的密集连接神经网络。这种方案很大程度上依赖于空间光调制器（SLM）来同时调节数百万个神经元的互联。然而，商用slm的主要缺点是速度和可集成性。高速、高亮度微型发光二极管（led）阵列集成互补金属氧化物半导体（CMOS）驱动电子器件的最新发展为高带宽和可集成性提供了难得的机会。特别是，在过去的几年中，我们的团队已经展示了具有ghz级调制带宽，亚微米像素间距和大像素计数的氮化镓oled阵列。因此，这种组合的uLED-on-CMOS阵列提供集成的、可重构的全光控制，消除了额外的电子调谐元件和外部光源的需要。在这个项目中，将探讨uLED-on-COMs阵列在元素数量和带宽方面的限制。此外，该系统将结合光学互连方案和先进的学习算法来构建集成光子神经网络。学生将获得大规模并行驱动电子（LED阵列驱动器）和PIC设计方面的专业知识，提供灵活的片上光电子集成。此外，学生将实现光子神经网络计算，并立即适用于复杂的任务，如全光信号再生和处理，光学模式识别和智能传感。这将需要获得自由空间光学设置和高级学习算法方面的专业知识。学生将成为一个更大的研究小组的一员，并有机会在一个充满热情的大学团队中与其他人一起工作。研究成果将在高影响力期刊上发表，并有机会在国际会议上发表。

Neuromorphic computing has gained huge momentum in the last decade thanks to the emergence of novel machine learning algorithms such as deep learning. Artificial neural networks are at the forefront of this revolution and their efficient hardware implementation poses significant challenges that impact on many fields of science and technology. The major problem posed by neural network computing is the handling of large matrix multiplications resulting from the parallel flux of information between densely connected layers of neurons. Photonics offers unique advantages for such demanding task as parallel operation is intrinsic to optical systems. In recent years, a number of schemes based on free-space optical setups have been proposed which have successfully implemented densely connected neural networks with hundreds of thousands of neurons. Such schemes largely rely on spatial-light modulators (SLM) to simultaneously tune millions of neuron interconnects. However, the main drawbacks of commercial SLMs are speed and integrability. An extraordinary opportunity for both high-bandwidth and integrability comes from the recent development of high-speed, high-brightness micro-light emitting diode (uLED) arrays integrated with complementary metal-oxide semiconductor (CMOS) drive electronics. In particular, gallium nitride uLED arrays with GHz-order modulation bandwidths, sub-micron pixel pitches, and large pixel counts have been demonstrated within the past few years by our group. As a result, such combined uLED-on-CMOS arrays offer integrated, reconfigurable, all-optical control eliminating the need of additional electronic tuning elements and external optical sources. In this project, the limits in terms of number of elements and bandwidth will be explored for uLED-on-COMs arrays. Moreover, this system will be combined with optical interconnectivity schemes and advanced learning algorithms to build integrated photonic neural networks. The student will gain expertise in massively parallel drive electronics (LED array drivers) and PIC design, delivering flexible photonic-electronic integration on-a-chip. Furthermore, the student will implement photonic neural network computing with immediate applicability to complex tasks like all-optical signal regeneration and processing, optical pattern recognition and smart sensing. This will require gaining expertise on free-space optical setups and advanced learning algorithms. The student will be part of a larger research group with the opportunity to work with others in a collegiate and enthusiastic team. Research findings will be published in high impact journals with the opportunity to present at international conferences.