Bio-inspired neuromorphic computing using memristor crossbar networks

Abstract:

In this thesis, we present experimental demonstrations of neuromorphic systems using fabricated memristor arrays as well as network-level simulation results. Models of resistive switching behavior in two types of memristor devices, conventional first-order and recently proposed second-order memristor devices will be first introduced. Secondly, experimental demonstration of K-means in a memristor network, an unsupervised clustering algorithm, will be presented. Thirdly, implementation of partial differential equation solver in memristor arrays will be discussed. This work expands the capability of memristor-based computing hardware from "soft' to "hard' computing tasks, which require very high precision and accurate solutions. In general first-order memristors are suitable to perform these tasks that are based on vector-matrix multiplications, ranging from K-means to PDE solvers. On the other hand, utilizing device dynamics in second-order memristors can allow natural emulation of biological behaviors and enable network functions such as natural encoding of temporal data. Our effort to explore second-order memristor device and its network behaviors will be discussed. Finally, analysis for large-size memristor arrays is presented, from the point of array fabrication and array operation based on experimental data and simulation.