A Study On Artificial Neuromorphic Devices And Neural Network Based On Novel High-density TaO_x Memristors

With the rapid development of the Microelectronic industry,the conventional CMOS technique is confronted with many challenges,such as the approaching of the physical limits and the extremely-high power density on integrated circuits,both of which prohibit its further scaling down process.Accordingly,in the past few decades,people are developing novel devices with new mechanisms,structures and materials,and novel computational architectures to remove the bottleneck.Inspired by the high computation efficiency of human brain,artificial neuromorphic devices and neuromorphic computing are believed to be promising candidates for future computing.Numerous novel neuromorphic devices have been proposed along with multiple novel neural network structures.However,these novel artificial neuromorphic devices either have drawbacks in integration or have deficiencies in faithfully mimicking biological functions,such as the synaptic plasticity,which limit its application for large scale,high accuracy neuromorphic computing.Hence,here we proposed a novel device structure,which is called the component graded TaO_x memristor.We discussed the design,fabrication of this novel device and systematically analyzed its significant improvements in DC performance and synaptic plasticity.Furthermore,focusing on the problems that the current hardware-based artificial neuromorphic networks are using simple network structure and dataset which cannot fully reveal the difficulty of real-life tasks and cannot provide enough hints for device design,we are using simulations to compare the influence of synaptic linearity,database complexity and the advanced extent of neural networks on the testing accuracy.We demonstrate the great importance of device linearity on improving testing accuracy.We also,for the first time,reveals the necessity of using complex database and network structure to evaluate device performance from the aspect of artificial neuromorphic network,in order to provide accurate instructions on device fabrications and designs.