E2CDA: Type II: Self-Adaptive Reservoir Computing with Spiking Neurons: Learning Algorithms and Processor Architectures

E2CDA：类型 II：带尖峰神经元的自适应储层计算：学习算法和处理器架构

• 批准号: 1639995
• 负责人: Peng Li
• 金额: $ 33.27万
• 依托单位: Texas A&M Engineering Experiment Station
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2016
• 资助国家: 美国
• 起止时间: 2016-09-01 至 2019-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1639995&HistoricalAwards=false
• 关键词: E2CDA; Type; II; Self; Adaptive

While computing has become increasingly data centric across many disciplines, conventional computer architectures have limited potential in meeting the escalating performance and energy efficiency needs in this era of data-driven science and engineering. This project aims to develop brain-inspired neural models of computation and adaptive processor architectures to enable intelligent data processing and learning in a wide range of applications. While being strongly interdisciplinary, this work will bridge neuroscience, artificial neural networks, computer architecture, and hardware engineering. The planned research will provide rich training and educational opportunities to students, and produce new curriculum. Research participation from undergraduate and underrepresented students will be promoted. The outcomes of this project will be broadly disseminated. Research collaboration with the US industry will be actively pursued via interaction with the Semiconductor Research Corporation. This work is aimed at attaining brain-like learning performance by imitating how the brain represents, processes, and learns from information, and more specifically, by developing models of computation based on the third-generation spiking neural networks, and efficient adaptive processor architectures. Within the framework of so called reservoir computing, the proposed neural models mimic key characteristics of the brain such as information processing based on spike timing. Furthermore, this project will develop brain-inspired learning mechanisms to allow training of complex recurrent spiking neural networks. Self-adaptive processor architectures with integrated on-chip learning, light-weight runtime learning performance prediction, and energy management will be developed to maximize system energy efficiency while providing a guarantee of performance.

虽然计算在许多学科中变得越来越以数据为中心，但传统的计算机架构在满足这个数据驱动的科学和工程时代不断升级的性能和能效需求方面的潜力有限。该项目旨在开发大脑启发的计算神经模型和自适应处理器架构，以在广泛的应用中实现智能数据处理和学习。虽然是强烈的跨学科，这项工作将桥梁神经科学，人工神经网络，计算机体系结构和硬件工程。计划中的研究将为学生提供丰富的培训和教育机会，并产生新的课程。将促进本科生和代表性不足的学生参与研究。该项目的成果将广泛传播。与美国工业界的研究合作将通过与半导体研究公司的互动积极进行。这项工作的目的是通过模仿大脑如何表示，处理和学习信息，更具体地说，通过开发基于第三代尖峰神经网络的计算模型和高效的自适应处理器架构来实现类似大脑的学习性能。 在所谓的水库计算框架内，拟议的神经模型模拟大脑的关键特征，例如基于尖峰定时的信息处理。此外，该项目还将开发大脑启发的学习机制，以训练复杂的递归尖峰神经网络。自适应处理器架构集成了片上学习，轻量级运行时学习性能预测和能源管理，将开发以最大限度地提高系统能源效率，同时提供性能保证。

项目成果

Online Adaptation and Energy Minimization for Hardware Recurrent Spiking Neural Networks

硬件循环尖峰神经网络的在线适应和能量最小化

• DOI: 10.1145/3145479
• 发表时间: 2018
• 期刊: ACM Journal on Emerging Technologies in Computing Systems
• 影响因子: 2.2
• 作者: Liu, Yu;Jin, Yingyezhe;Li, Peng
• 通讯作者: Li, Peng

Exploring sparsity of firing activities and clock gating for energy-efficient recurrent spiking neural processors

探索节能循环尖峰神经处理器的发射活动和时钟门控的稀疏性

• 发表时间: 2017
• 期刊: Proceedings - International Symposium on Low Power Electronics and Design
• 作者: Liu, Y;Jin, Y
• 通讯作者: Jin, Y

Calcium-modulated supervised spike-timing-dependent plasticity for readout training and sparsification of the liquid state machine

用于读出训练和液态机稀疏化的钙调制监督尖峰时间依赖性可塑性

• 发表时间: 2017
• 期刊: Proceedings of ... International Joint Conference on Neural Networks
• 作者: Jin, Y