忆阻神经形态计算系统的理论分析与电路设计

The implementation of memristor provides a novel way to design circuits with new functions. Based on the specific properties of memristor, such as nonlinear dynamical response characteristics which is more in line with the cognitive computing of neural networks than the general electronic components, this project aims to resolve several fundamental challenges in the research and development of memristor crossbar based neuromorphic computing system (MC-NCS) designs via: establishing the advanced theory required to reliably and efficiently perform neuromorphic computation on memristor crossbars, taking into account current ideas surrounding spiking-based computation, memristor-based spatio-temporal (MST) synapse design, computing workload partitioning, and etc.; creating modeling methodologies, as well as the relevant optimizations, required to accurately simulate the electrical properties of memristor devices, reducing the computing complexity of MC-NCS while enhancing training reliability; and developing circuit, physical and architecture design techniques for robust and scalable MC-NCS design.The research outcomes will benefit the computer system community through the development of robust, scalable, and interdisciplinary design methodologies allowing the construction of neuromorphic computing circuits utilizing emerging memristor technology. These technical findings, together with trained human resources, will fulfill national critical technology needs and greatly enhance the relevant research infrastructure.

忆阻的成功研制为电子电路的设计提供了新的途径，并使该电路具有新的功能。本项目利用忆阻特有的，符合神经网络认知计算功能的非线性动态响应特性，提出基于忆阻交叉阵列结构的可靠和可扩展的计算理论，实现将具有代表性的大规模的神经网络计算能力应用到具有独特电学性质的忆阻交叉阵列中去；考虑到阵列压降存在对忆阻阵列的可编程性和计算精度的影响，提出能够准确模拟忆阻器和忆阻交叉阵列的电气性能的综合建模和优化方法，减少计算复杂度，提高运行可靠性；为提高交叉阵列的使用，减少阵列面积和功耗，开发相关的分层物理架构体系，来降低系统的复杂性，并提高鲁棒性和可扩展性，使神经形态计算在异构系统上加速运行。通过鲁棒性、可扩展性、和跨领域设计理论的研究，促进忆阻器件模型、神经形态计算算法和系统实现等方面的经验累积和理论创新。这些技术成果连同训练有素的人力资源，不仅满足当前的关键技术需要，也会极大提高相关领域的研究水平。

利用忆阻神经形态计算在大规模数据处理方面显示出了巨大的潜力，实现单一设备或小型网络上的突触行为等已经完成不能够满足未来应用神经形态计算实现大数据挖掘等方面发展趋势的需要。根据前期工作以及忆阻交叉阵列在神经形态计算的巨大潜力，将利用忆阻交叉阵列，研究基于忆阻交叉阵列的神经形态计算系统的先进理论，实现低能耗和强大计算能力的计算系统，提出相应的系统建模与设计方法，提高系统的鲁棒性和可扩展性。建立新的忆阻神经形态计算基本理论体系以适应大规模、可扩展的网络所引起的技术挑战，缩小忆阻理论研究与集成计算系统的高性能、低成本、节能等商业化需求之间的差距。通过鲁棒性、可扩展、跨领域设计理论的研究，促进计算机系统领域发展，以及基于现有忆阻技术的神经形态计算电路结构设计，引领忆阻器件模型、神经形态计算算法和系统实现等方面的经验累积和理论创新。这些技术成果连同训练有素的人力资源，不仅满足当前的关键技术需要，也会极大提高相关领域的研究水平。