CRI: CI-NEW: Trainable Reconfigurable Development Platform for Large-Scale Neuromorphic Cognitive Computing

神经形态认知计算旨在通过模拟生物大脑中高效和有弹性的自适应信息处理的原理和物理组织来学习解决复杂的认知任务。尽管经过了30多年的发展，最近所有科学、技术、工程和数学(STEM)学科的广泛兴趣激增，但由于目前现有系统的专业性和复杂操作的高准入门槛和成本，神经形态认知计算仍然主要局限于该领域训练有素的研究人员的一小部分。该项目将构建和支持一个通用的神经形态认知计算平台，它将是迄今为止实现的最大和最多功能的平台，也是第一个广泛可用并向广大研究界开放的平台，用于研究更有效和更高效地接近人类大脑认知能力的新形式的大脑启发计算。该平台的目标是在更广泛的STEM研究社区中被不同领域的用户广泛采用，该平台将具有一个自然的用户界面，通过提供一套由用户社区维护并与用户社区共享的用户友好软件工具，使新手用户免受操作和配置高度专业化的神经形态硬件所产生的挑战。在圣地亚哥超级计算机中心用于用户访问和数据共享的广泛现有网络和存储基础设施的基础上，该平台将通过神经科学网关(NSG)门户网站进行托管和维护，该门户网站目前为科学界超过600名活跃用户提供服务。大型神经形态平台将作为计算机和信息科学与工程(CEISE)研究界无与伦比的新资源，满足对实验试验台的巨大需求，以研究超越传统冯·诺伊曼范式的替代计算形式，以及计算技术扩展中摩尔定律扩展即将面临的物理限制。可重新配置的平台将以内存计算处理节点的分层互连网络为特色，该网络实时模拟多达1.28亿个神经元的高度灵活的神经动力学(集成和开火、分级、随机二进制等)，以及多达320亿个突触的高度灵活的连接性和可塑性(尖峰时间依赖的可塑性、基于梯度的深度学习等)。该系统将能够在计算神经科学建模中提供生物物理细节，以及在线自适应模式识别方面的高性能和高效率，为传统上追求不同计算方法的计算神经科学和计算智能社区提供服务并将其聚集在一起。该平台的用户界面将支持软件工具和资源，用于人工智能应用中的深度学习和运行时优化，以及计算神经科学研究中记录的神经活动对结构和功能连接的干扰等。为了促进最大的科学和社会影响，基础设施将在时间管理的共享基础上免费提供给任何研究人员，以换取同意共享复制文献中报告的结果所需的源代码和数据。该奖项反映了NSF的法定使命，并通过使用基金会的智力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 cognitive computing aims at learning to solve complex cognitive tasks by emulating the principles and physical organization of highly efficient and resilient adaptive information processing in the biological brain. Despite over 30 years of development and a recent surge of broad interest across all Science, Technology, Engineering and Mathematics (STEM) disciplines, access to neuromorphic cognitive computing remains mostly limited to a small community of highly trained researchers in the field due to high entry barriers and costs associated with the specialized nature and complex operation of currently available systems. This project will construct and support a general-purpose neuromorphic cognitive computing platform that will be the largest and most versatile realized to date as well as the first to be broadly available and open to the research community at large, for research into new forms of brain-inspired computing that are more effective and more efficient in approaching the cognitive capabilities of the human mind. Targeting wide adoption by a diverse cross-section of users in the broader STEM research community, the platform will feature a natural user interface that shields novice users from the challenges arising in operating and configuring highly specialized neuromorphic hardware, by providing a set of user-friendly software tools maintained by and shared with the user community. Building on extensive existing network and storage infrastructure for user access and data sharing at the San Diego Supercomputer Center, the platform will be hosted and maintained through the Neuroscience Gateway (NSG) Portal, which currently serves over 600 active users in the scientific community.The large-scale neuromorphic platform will serve as a new and unparalleled resource to the Computer and Information Science and Engineering (CISE) research community, addressing a great need for an experimental testbed for research in alternative forms of computing beyond the traditional von Neumann paradigm and the impending physical limits to Moore's Law expansion in the scaling of computing technology. The reconfigurable platform will feature a hierarchically interconnected network of in-memory computing processing nodes that emulates, in real-time, highly flexible neural dynamics (integrate-and-fire, graded, stochastic binary, etc) of up to 128 million neurons with high flexible connectivity and plasticity (spike-timing dependent plasticity, gradient-based deep learning, etc) of up to 32 billion synapses. The system will be capable of biophysical detail in computational neuroscience modeling, as well as high performance and efficiency in on-line adaptive pattern recognition, serving and bringing together both computational neuroscience and computational intelligence communities that have traditionally pursued disparate computational approaches. The user interface of the platform will support software tools and resources for deep learning and run-time optimization in artificial intelligence applications, and for interference of structure and functional connectivity from recorded neural activity in computational neuroscience research, among others. To facilitate greatest scientific and societal impact, the infrastructure will be made available free of charge, on a time-managed shared basis, to any researcher in return for agreeing to share source code and data necessary to replicate results reported in the literature.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.