Neuromorphic Computing Application Based On Memristors

In the near future,the development of Internet of Things and Artificial Intelligence need much more computing sources to process the big data,which pose a big challenge for traditional computing systems based on CMOS technology.In the device level,Moore's law is coming to the end—as we try to continuously scale down the device size to 10 nm,it's hard to cut off the current between the source terminal and drain terminal in CMOS FET due to the quantum tunneling effect,which would make the failure in logical computing.In the architecture level,the most popular paradigm of computing system dominating the market is famous von Neumann computer,which involves two separated part of processing unit and memory.The heavy task of data communication between the processing unit and memory—known as von Neumann bottleneck—leads to poor real-world performance and high power consumption.Those limits bring researchers to new devices and new architecture to realize more powerful,more energy-efficient computing application even replacing the traditional computing systems in some field.Especially in recent years,the popularization of artificial neural network is speeding up this trend because the traditional computing systems get hard to meet the demand in running large-scale neural network tasks.Although some modified methods such as GPU,TPU and NPU have been proposed to optimize the traditional computing systems,energy and time issues still need to be solved with a revolutionary way.One of promising solutions is neuromorphic computing,namely brain-inspired computing.The researchers hope this revolutionary method to processing information high efficiently in a way as human brain works.In this paper,we firstly introduced the traditional computer science and artificial neural network.The challenge of traditional computing systems as well as the need for more efficient computing system motivate the interest in neuromorphic computing.As one of candidates,memristor is a promising device to develop the neuromorphic computing to a new era.In Chapter two,we introduce the switching mechanism in theory,and illustrate the fabrication of memristor and memristor crossbar array.Subsequently,we show the characterization methods and summary of electrical performance of our fabricated memristors in Chapter three.And then in Chapter four,we assembled a robotic vehicle,which used a memristive neuromorphic circuit as its processor to process the information from sensors and control the motor to react to the changing environment.The plasticity of memristor similar with biological synapse enables the changeable behaviors of the vehicle by training the memristive neuromorphic processor.A supervised learning process allow the vehicle to perform the tasks of line-tracking and obstacle-avoiding.Moreover,because the information processing is implemented in the analog domain without transformation between analog and digital domains,the response latency of the memristive neuromorphic processor is only 56 ns,which is around 700 x faster than a MCU chip.In next Chapter five,we research a novel reconfigurable FET,which is hoped to build a novel 1T1 R memristor array controled by a single global gate.This 1T1 R array is promising due to the simple control and high performance.We finally conclude the paper in Chapter six and give the outlook for the neuromorphic computing.From my perspective,these important challenges are essential to be further researched: first,large-scale multiple layer neuromorphic computing;second,intelligent neuromorphic robotics using multiple sensors.Maybe my future research interest will be concentrated in these points.