Expeditions: Coherent Ising Machines for Optimization, Machine Learning and Neuromorphic Computing

这个探险项目旨在更深入地了解相干伊辛机（CIM）的基本性质和潜在用途。这些机器利用非传统的计算架构来解决从物流和机器人到材料工程和药物设计等应用领域的关键优化问题。 基于当前CIM原型的性能，下一代CIM有望在这些领域推动人工智能（AI）能力的实质性实际进步。 CIM也是重要的基础研究兴趣，作为新的架构，我们可以在后摩尔定律时代测试计算机工程的变革思想。 CIM利用光学和电子元件的协同组合来实现大规模数据连接和快速可编程逻辑。 他们同样利用了一种非传统的光学存储器，这是迈向更激进的量子信息技术的垫脚石。 CIM原型今天实现了更大的规模，同时利用了比原型量子计算机更多的设备物理进步。 因此，开发新的CIM应用程序，对CIM原型进行基准测试，并分析CIM缩放，可以为传统技术和理想化量子机制之间的计算物理学的知之甚少的方面提供新的启示。 CIM依赖于数据处理原语，与深度学习所需的数据处理原语有很大的相似之处，这表明CIM研究所产生的技术创新可能会对人工智能的硬件开发产生更广泛的影响。 因此，这次CIM远征的核心研究将成为量子和神经形态计算正在进行的努力的重要补充。 这次CIM远征将支持先进原型硬件的开发，包括纳米光子学，光电子学和超快激光源的最新进展。 在继续探索光学和电子技术在专业优化/人工智能架构中的互补性时，其研究人员将特别感兴趣的是评估原始速度和能源效率之间的权衡，这些权衡可以在这种混合设计空间中以新的方式进行。 该项目将利用水库计算和消息传递算法的见解研究CIM架构的泛化，并建立量子效应在CIM中的作用的一般理论，包括稳健地利用它们的策略。 该项目将扩展最近对深度学习神经网络动力学的理论分析，以阐明硬优化问题的数学结构与物理计算机器的解决方案轨迹之间的微妙联系。 CIM Expedition团队将与工业合作伙伴和应用领域专家合作，对现有的CIM原型进行广泛的基准测试，预测可行的CIM性能，并与传统计算方法以及新兴量子计算进行比较。 这项工作将遵循最佳实践，以便在不同的硬件平台和组合优化和机器学习中的实际应用用例之间进行比较，这是由CIM Expedition团队的成员先前建立的。 总的来说，该项目的目标是更好地理解当前CIM原型的功能和操作原理，以及更清楚地了解CIM扩展和基本性能改进的未来可能性。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This Expeditions project seeks to gain a deeper understanding of the fundamental nature and potential uses of Coherent Ising Machines (CIMs). These machines exploit unconventional computing architectures to solve crucial optimization problems for application domains ranging from logistics and robotics to materials engineering and drug design. Based on the performance of current CIM prototypes, next-generation CIMs hold great promise to drive substantial practical advances in artificial intelligence (AI) capabilities in such fields. CIMs are of significant fundamental research interest as well, as novel architectures with which we can test transformative ideas for computer engineering in the post-Moore’s Law era. CIMs exploit a synergistic combination of optical and electronic components to achieve both massive data connectivity and fast programmable logic. They likewise utilize an unconventional optical memory that represents a stepping stone towards more radical quantum information technologies. CIM prototypes today achieve substantially greater scale while leveraging more incremental advances in device physics than do prototype quantum computers. Work to develop new CIM applications, benchmark CIM prototypes, and analyze CIM scaling can thus shed new light on poorly understood aspects of the physics of computation that sit between conventional technology and an idealized quantum regime. CIMs rely upon data processing primitives with substantial parallels to those required for deep learning, suggesting that technical innovations arising from CIM research may have broader impact in hardware development for AI. The core research of this CIM Expedition will thus serve as an important complement to ongoing efforts towards quantum and neuromorphic computing. This CIM Expedition will support the development of advanced prototype hardware incorporating recent advances in nanophotonics, optoelectronics and ultrafast laser sources. In continuing to explore the complementarity of optics and electronics for specialized optimization/AI architectures, its researchers will be particularly interested in assessing tradeoffs between raw speed and energy efficiency that can be made in new ways in this hybrid design space. The project will investigate generalizations of CIM architectures utilizing insights from reservoir computing and message passing algorithms, and establish a general theory of the role of quantum effects in CIM including strategies to exploit them robustly. The project will extend recent theoretical analyses of the dynamics of deep learning neural networks to elucidate subtle connections between the mathematical structure of hard optimization problems and the solution trajectories of physical computing machines. The CIM Expedition team will work with industrial partners and applications-domain specialists to perform extensive benchmarking of existing CIM prototypes, forecasting feasible CIM performance in comparison with conventional computing approaches as well as emergent quantum computing. This work will follow best practices for making comparisons across disparate hardware platforms and realistic application use cases in combinatorial optimization and machine learning, established previously by members of the CIM Expedition team. Overall the project will aim to develop a sharper understanding of the capabilities and operating principles of current CIM prototypes, as well as a clearer picture of future possibilities for CIM scaling and fundamental performance improvements.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.