Neuromorphic Computing with Memristors: From Devices to Integrated Systems

Abstract

Neuromorphic computing is a concept to use electronic analog circuits to mimic neuro-biological architectures present in the nervous system. It is designed by following the operation principles of human or mammal brains and aims to use analog circuits to solve problems that are cumbersome to solve by digital computation. Neuromorphic computing systems can potentially offer orders of magnitude better power efficiency compared to conventional digital systems, and have attracted much interest recently. In particular, memristors and memristor crossbar arrays have been widely studied for neuromorphic and other in-memory computing applications. Memristors offer co-located memory and logic functions, and intrinsic analog switching behaviors that enable online learning, while memristor crossbars provide high density and large connectivity that can lead to high degree of parallelism. This thesis work explores the device characteristics and internal dynamics of different types of memristor devices, as well as the crossbar array structure and directly integrated hybrid memristor/mixed-signal CMOS circuits for neuromorphic computing applications. WOx-based memristors are used throughout the thesis. Bipolar resistive switching is observed due to oxygen vacancy redistribution within the switching layer upon the application of an applied electric field. In a typical WOx memristor, oxygen vacancy drift by electric field and spontaneous diffusion result in a gradual resistance change. Depending on the purpose of the applications, the oxidation condition can be varied to achieve either short-term memory or long retention properties, which allow the devices to be used in applications such as reservoir computing or learning and inference. Device fabrication and modeling are briefly discussed. A network structure can be directly mapped onto a memristor crossbar array structure, with one device formed at each crosspoint. When an input vector is fed to the network (typically in the form of voltage pulses), the output vector can be obtained in a single read process, where the input-weight vector-matrix multiplication operation is performed natively in physics through Ohm’s law and Kirchhoff’s current law. This elegant approach of implementing matrix operations with memristor network can be applied for many machine learning algorithms. Specifically, we demonstrate a sparse coding algorithm implemented in a memristor crossbar-based hardware system, with results applied to natural image processing. The system is also estimated to achieve ~16× energy efficiency than conventional CMOS system in video processing. We further fabricated a 54×108 passive memristor crossbar array directly integrated with all necessary interface circuitry, digital buses and an OpenRISC processor to form a complete hardware system for neuromorphic computing applications. With the fully-integrated, reprogrammable chip, we demonstrated multiple models such as perceptron learning, principal component analysis, and also sparse coding, all in one single chip, with power efficiency of 1.3TOPS/W. The internal device dynamics, including the short-term memory effect caused by spontaneous oxygen vacancy diffusion, additionally allows us to implement a reservoir computing system to process temporal information. Tasks such as handwritten digit recognition are achieved by converting the spatial information of a digit image into streaming inputs fed into a reservoir composed of memristor devices. The system is also used to experimentally solve a second-order nonlinear task, and can successfully predict the expected output without knowing the original dynamic transfer function. Other attempts to explore the potential of using memristor networks to solve challenging problems more efficiently are also investigated. Two typical problems, including Hopfield network and self-organizing maps will be discussed.