REservoir COMputing with MEmristive Nonlinear Dynamics: Theory, Design and Applications (RECOMMEND)

使用忆阻非线性动力学进行储层计算：理论、设计和应用（推荐）

• 批准号: 536063366
• 负责人: Professor Dr. Fernando Corinto, Ph.D.
• 金额: --
• 依托单位: Professur für Grundlagen der Elektrotechnik
• 依托单位国家: 德国
• 项目类别: Priority Programmes
• 资助国家: 德国
• 项目状态: 未结题
• 来源: https://gepris.dfg.de/gepris/projekt/536063366?language=en
• 关键词: REservoir; COMputing; MEmristive; Nonlinear; Dynamics

Reservoir computing (RC) is an efficient machine learning method for temporal/sequential data processing. RC uses a nonlinear representation of the input data in a high-dimensional space called a reservoir. Recently, this method has been shown to have similarities to statistical methods such as nonlinear vector autoregression, which does not require a high-dimensional reservoir and therefore can be implemented with a smaller number of components and produces interpretable results. The Next Generation Reservoir Computing (NGRC) concept combines these approaches and promises extremely efficient neuromorphic circuit implementations. The goal of the REservoir COMputing with MEmristor Nonlinear Dynamics: Theory, Design and Applications (RECOMMEND) project is to realize a reconfigurable, energy-efficient neuromorphic computing system at the hardware and modeling levels by using theoretical foundations to develop a hardware prototype of a versatile RC platform based on 1) a low-dimensional nonlinear dynamic reservoir circuit with memristors as nonlinear tunable dynamic elements. 2) computational/memory elements that enable one-step in-memory vector-matrix multiplication (VMM). 3) supervised control circuitry that enables multitasking and correct functionality under time-varying conditions. Hardware versions of memristors that exhibit electrically induced resistance change, along with scalable electronic components (resistors, capacitors), will form the analog core of the RC platform. Using circuit and system theory principles (flux charge analysis and Volterra methods), the resulting tunable nonlinear dynamics of the devices will be used to map the input data into an appropriate vector of nonlinear elements to control the output layer of the network for time series prediction and optimization problem solving. Research is conducted at the intersection of several complementary technical fields, including materials science, device physics, and circuit and system theoretical modeling, including analytical methods to support circuit design, analysis of nonlinear dynamic characteristics of nonautonomous circuits, to the development of mathematical tools and algorithms. The goal of our multidisciplinary consortium is thus to enable a radical paradigm shift in information processing towards energy-efficient, robust, and scalable RC platforms with memristors that bridge the gap between statistical methods and machine learning.

水库计算（RC）是一种有效的机器学习方法的时间/顺序数据处理。RC使用输入数据在称为库的高维空间中的非线性表示。最近，这种方法已被证明具有类似的统计方法，如非线性向量自回归，它不需要一个高维的水库，因此可以实现较少数量的组件，并产生可解释的结果。下一代水库计算（NGRC）的概念结合了这些方法，并承诺非常有效的神经形态电路实现。RESERVOIR COMputing with MEmristor Nonlinear Dynamics：Theory，Design and Applications（RECOMMEND）项目的目标是在硬件和建模层面上实现可重构的、节能的神经形态计算系统，方法是利用理论基础开发一个通用RC平台的硬件原型，该平台基于1）以忆阻器作为非线性可调动态元件的低维非线性动态水库电路。2)计算/存储器元件，其实现一步式存储器内向量矩阵乘法（VMM）。3)受监督的控制电路，能够在时变条件下进行多任务处理并实现正确的功能。表现出电致电阻变化的忆阻器的硬件版本，沿着可扩展的电子元件（电阻器、电容器），将形成RC平台的模拟核心。使用电路和系统理论原理（通量电荷分析和沃尔泰拉方法），所得到的装置的可调非线性动态将用于将输入数据映射到非线性元件的适当向量中，以控制网络的输出层，用于时间序列预测和优化问题求解。研究在几个互补的技术领域的交叉点进行，包括材料科学，器件物理学，电路和系统理论建模，包括支持电路设计的分析方法，非自治电路的非线性动态特性分析，数学工具和算法的开发。因此，我们的多学科联盟的目标是实现信息处理的根本范式转变，转向节能，鲁棒和可扩展的RC平台，其忆阻器弥合了统计方法和机器学习之间的差距。