Application Of Memristive Neuromorphic System In Pattern Recognition

Neuromorphic system is a hardware implementation of artificial neural network with higher capacity of information processing and fault tolerance.It is extensively used in pattern recognition,machine learning,signal processing and image processing.Memristor acts as a nano-scale device with nonvolatile/volatile,memory,plasticity,low power consumption and can serve as a natural synapse.The memristive crossbar array can be utilized as a natural weight matrix in neuromorphic system.Because of the different characteristics in different memristor devices,neuromorphic systems based on various memristor devices are becoming abundant in order to meet different requirements.But most of these systems have three issues.First,test samples are largely preprocessed by traditional method.Second,training process of network is usually realized by off-line system,and testing process is realized in neuromorphic circuit system.Third,most of the neuromorphic systems built on the traditional digital system approaches which ignore the unique characteristics of the human brain,such as forgetting effect.In order to solve these issues,this thesis presents a variancecorrelation algorithm based forgetting memristive neuromorphic system and it is successfully applied to pattern recognition.In addition to the effective recognition of handwritten number images,the relationship between the forgetting rate and the recognition rate is also discussed.Specifically,the main research results in this thesis are as follows.Firstly,the internal mechanisms of classical HP memristor and forgetting memristor are described,and their mathematical models are deduced.Based on the one-dimensional forgetting memristor model,the improved three-dimensional model is introduced.And unipolar,bipolar and reversible conditions of the three-dimensional model are given.The characteristics and synaptic behavior of these three memristor models are compared and analyzed in detail by SPICE model simulations.Second,a forgetting memristor based neuromorphic circuit system is presented.It is a multi-layer integrated system which contains a self-learning circuit,a training circuit and a recognition circuit,and it can accomplish training and recognition on-line.In view of the different functions of different layers in system,the structural circuit designs and function simulations of each part are given.Third,based on the group characteristics and individual characteristics of samples,a variance-correlation learning algorithm is proposed to realize the on-line training of the memristor crossbar array.This method can simplify the preprocessing of the samples,while facilitating the implementation of circuit system.Finally,the training and recognition of handwritten number images are regarded as an application for this system and the simulations verify its effectiveness.In addition,the influence of forgetting factor on recognition results is further discussed.Simulations indicate that different value intervals of forgetting factors have different level effects on the recognition rate.