FET: Small: Heterogeneous Learning Architectures and Training Algorithms for Hardware Accelerated Deep Spiking Neural Computation

该项目旨在通过开发高能效的新尖峰神经架构、训练算法和硬件计算设备，解决当前数据驱动应用程序广泛领域中的计算性能和能效危机。来自生物大脑的灵感将被用来支持算法和硬件系统的发展，以缩小计算能力供需之间日益扩大的差距。该项目的结果将是强烈的跨学科，预计将刺激机器学习的技术进步，以及神经网络、神经科学和硬件工程之间的桥梁。该研究将为学生提供丰富的培训和教育机会。将通过各种外展计划促进本科生和代表性不足群体的研究参与。该项目的成果将在广泛的研究和工业界传播，并纳入研究生课程。将寻求与业界的研究合作，以指导这项工作，以应对现实世界的挑战，并在工业环境中为学生提供指导和培训。受大脑启发的计算模型和硬件计算系统有望提供在后摩尔定律时代处理日益庞大的数据所需的计算能力，而不会产生相应的高能源成本。该项目将致力于通过解决两个紧迫的相互依赖的研究障碍来提高棘波神经模型在现实生活学习任务中的性能：缺乏计算能力强大的学习架构，以及缺乏能够有效训练复杂的棘波神经模型的实用算法。将探索神经科学和深度学习之间的协同作用，以开发不同类型的深度棘波神经体系结构和学习算法，以解决相应的训练瓶颈。高效的尖峰神经处理器将在可重新配置的计算设备上进行演示。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This project aims to address the present performance and energy efficiency crisis in computing across broad areas of data-driven applications by developing energy-efficient new spiking neural architectures, training algorithms, and hardware computing devices. Inspirations from biological brains will be taken to support the development of algorithms and hardware systems to close the widening gap between the supply and demand of computing power. The outcomes from this project will be strongly interdisciplinary and are expected to stimulate technical advancements in machine learning and bridge between neural networks, neuroscience, and hardware engineering. The research will provide rich training and educational opportunities to students. Research participation from undergraduate students and underrepresented groups will be promoted through various outreach programs. The results of this project will be disseminated in broad research and industrial communities and integrated into the graduate-level curriculum. Research collaboration with industry will be sought to guide this work toward addressing real-world challenges and provide mentoring and training of students in the industrial setting. Brain-inspired models of computation and hardware computing systems hold the promise of delivering the amount of computing power required in processing increasingly large volumes of data in the post Moore's Law era, without a correspondingly high energy cost. This project will focus on improving the performances of spiking neural models for real-life learning tasks by addressing two pressing inter-dependent research roadblocks: lack of computationally powerful learning architectures, and lack of practical algorithms that can effectively train complex spiking neural models. Synergies between neuroscience and deep learning will be explored to develop heterogeneous deep spiking neural architectures and learning algorithms to address the corresponding training bottlenecks. Efficient spiking neural processors will be demonstrated on reconfigurable computing devices.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.