EFRI BRAID: Rapid contextual learning in resilient autonomous systems

神经形态计算旨在识别大脑的关键操作原理，并将其应用于人工计算系统中。这项工作包括基于去中心化大脑网络架构的新硬件平台（如英特尔Loihi和IBM TrueNorth平台），以及有效运行新硬件所需的新兴计算算法。在硬件方面，这种新计算范式的诊断原则是并行、异步本地计算以及内存和计算资源的共定位。这意味着成千上万的小处理器都是单独运行的——不使用公共时钟、共享内存存储或任何其他公共资源，这些资源会使整个系统的速度降低到最慢的组件的速度。其结果是，计算机系统可以更快地执行许多类型的任务，并且能耗更低，但这需要对软件算法进行彻底的重新思考，以便使用这些基本上分散的电路有效地执行现实世界的任务。在本应用中，从生物脑回路中提取的计算原理被用于开发这样的工作算法，并且还用于识别和分析这些算法中的核心计算基序，以便将来重新利用。从神经科学中提取的其他原理也将被实施和评估，特别是局部复杂性和异质性，其中“神经元”可以单独复杂并且彼此非常不同，以及自适应网络扩展，其中网络本身可以根据其获得的学习和专业知识增长。康奈尔大学和佐治亚理工学院将通过多项举措扩大对这些变革性计算战略的培训和接触，从历史上成功的多样性、公平性和包容性战略，到K-12合作伙伴关系，再到沉浸式STEM教学设施和外展项目。神经形态计算平台的潜在优势既清晰又深刻，但也受到缺乏成熟的神经形态算法的限制，这些算法能够利用这些优势来解决现实世界的问题。Sapinet网络基于从生物嗅觉系统中提取的计算原理，有望成为一种用于噪声下信号恢复和识别的神经形态算法。使用明确的理论路线图，将开发该网络架构，以结合额外的大脑启发策略，以实现弹性和强大的自主性，例如上下文依赖、多模态集成、丰富类别学习和明确的相似性表示，这些策略共同承诺实现更优越和更复杂的性能。其次，部分由于构成其能力的设计元素的异质性，神经形态计算目前受到缺乏形式化分析和优化技术的限制。将开发一套神经形态操作的计算基序（“数值配方”）和分析策略，以服务于系统神经科学中可能缺乏明确并行的未来应用。最后，由此产生的智能系统将在软件和神经形态硬件中实例化，并最终在实际部署和测试的原型设备中实例化。总体目标是构建和部署具有高度性能自主性的本地智能、节能和便携式边缘设备；即，在次优和不可预测的现实世界条件下表现出弹性和上下文感知的任务性能。该奖项反映了美国国家科学基金会的法定使命，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。

Neuromorphic computing seeks to identify key operational principles of the brain and implement them in artificial computing systems. The effort comprises both new hardware platforms with architectures based on decentralized brain networks (such as the Intel Loihi and IBM TrueNorth platforms) and the emerging computational algorithms that are required to run this new hardware effectively. In hardware, the diagnostic principles of this new computing paradigm are parallel, asynchronous local computation and the colocalization of memory and compute resources. This means that thousands of small processors all operate separately – without using a common clock, a shared memory store, or any other common resources that would slow the whole system down to the speed of its slowest component. The result is a computer system that can perform many types of tasks much faster, and with much lower energy expenditure, but that requires a complete rethinking of software algorithms in order to perform real-world tasks effectively using these fundamentally decentralized circuits. In the present application, computational principles extracted from biological brain circuits are employed to develop such working algorithms, and also to identify and analyze core computational motifs from these algorithms for future repurposing. Additional principles drawn from neuroscience also will be implemented and assessed, particularly local complexity and heterogeneity, in which the “neurons” can be individually complex and very different from one another, and adaptive network expansion, in which the network itself can grow in accordance with its acquired learning and expertise. Training and exposure to these transformative compute strategies will be broadened via multiple initiatives at Cornell and Georgia Tech, ranging from historically successful diversity, equity, and inclusion strategies to K-12 partnerships to immersive STEM teaching facilities and outreach programs. The potential advantages of neuromorphic computing platforms are both clear and profound, but also are limited by the paucity of well-developed neuromorphic algorithms capable of leveraging these advantages to address real-world problems. The Sapinet network, based on computational principles extracted from the biological olfactory system, shows promise as a neuromorphic algorithm for signal restoration and identification under noise. Using an explicit theoretical roadmap, this network architecture will be developed to incorporate additional brain-inspired strategies for resilient and robust autonomy, such as context dependence, multimodal integration, rich category learning, and explicit representations of similarity that together promise to enable superior and more sophisticated performance. Second, owing in part to the heterogeneity of design elements that underlie its power, neuromorphic computing presently is limited by a paucity of formal analysis and optimization techniques. A set of computational motifs (“numerical recipes”) and analysis strategies for neuromorphic operations will be developed, in service to future applications that may lack an explicit parallel in systems neuroscience. Finally, the resulting intelligent systems will be instantiated in software and in neuromorphic hardware, and ultimately in prototype devices for real-world deployment and testing. The overall goal is to construct and deploy locally intelligent, energy-efficient, and portable edge devices capable of a high degree of performance autonomy; i.e., that exhibit resilient and context-aware task performance under suboptimal and unpredictable real-world conditions.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.