局部有源忆阻器的神经形态动力学

Neuromorphic Dynamics of Locally Active Memristors

在算力需求指数增长的驱动下,计算机芯片晶体管密度的微缩接近物理极限,传统存算分离的冯诺依曼计算架构已形成速度和功耗瓶颈,一种有希望的替代方案是基于忆阻器神经形态动力学的类脑计算。忆阻器能够模拟神经元和突触行为在物理层面直接处理与存储信息,可实现高效能、高算力和存算一体的类脑计算新架构。在文中主要介绍作者团队在局部有源忆阻器神经形态动力学研究方面的最新进展,其中包括局部有源忆阻器特性及建模,基于混沌边缘理论和局部有源忆阻器的人工神经元模型及其设计方法,忆阻神经元动作电位的动力学机理,对Smale悖论的理论解释和忆阻神经网络中的奇美拉态,以及忆阻神经元的物理实现及实验等。

Driven by exponential growth of computing power demand,the miniaturization of transistor density in computer chips is approaching the physical limit,and the traditional Von Neumann computing architecture,which separates storage and computing,has formed speed and power bottlenecks.A promising alternative scheme is brain-like computing based on memristor neuromorphic dynamics.Memristors can imitate the behaviors of neurons and synapses to directly process and store information at the physical level,achieving a new brain-like computing architecture with high energy efficiency,high computing power and in-memory computing.In this review paper,the latest progress in the study on neuromorphic dynamics of local active memristors is mainly introduced,including the characteristics and modeling of locally active memristors,artificial neuron models and design methods based on edge of chaos theory and locally active memristors,dynamic mechanisms of memristor-based neuron action potentials,theoretical explanations of the Smale paradox and the Chimera state in memristive neural networks,as well as the physical implementations and experiments of memristor-based neurons.