Memristor Based Circuit Analysis And Its Applications In Neuromorphic Systems

With the continuous development of computer and electronic information technology,greater demands have been placed on the information acquisition mode,transmission speed,storage capacity,and processing efficiency.Since the Von Neumann architecture-based calculation pattern lacks of adaptability(i.e.,the self-learning and evolution cannot be developed during the calculation process),some popular calculation systems have encountered bottleneck in intellisense,pattern recognition,decision control,etc.As a kind of intelligent information processing systems vaguely inspired by the biological neural system,the artificial neural networks(ANNs)possess numerous properties,such as the parallel-distributed processing capability,the nonlinear processing capability,and the self-learning capability.Currently,the relevant work mainly focuses on the theoretical analysis of neural network construction and its learning algorithms.The corresponding neural network hardware implementation(i.e.,the neuromorphic system)is still in the early stage of development.Traditional neuromorphic system realization schemes always failed in dealing with the trade-off among the processing speed,accuracy,and occupied area,due to some issues such as the circuit area,power consumption,compatibility,integration degree,and so forth.The recent advent of memristor has opened up a new approach for addressing the aforementioned issues occurred in the existing neuromorphic systems,on account of its advantages of nonvolatility,variable conductivity,threshold feature,and nanoscale geometries.So far,a variety of memristor-based neuromorphic systems have been put forward.However,most of them have been suffering from three basic issues:1)The multiple memristor composite circuits,as the fundamental elements in memristor based neuromorphic systems have not been discussed in detail;2)The memristive synapses always lack completeness in weight sign,and the linearity in weight processing and weight programming cannot be easily achieved;3)The inherent parameter estimation problem cannot be addressed effectively.Hence,based on the recent research achievement in modern neuroscience,electronic circuit theory,and digital image processing,this work mainly investigates the remedies of the abovementioned issues.The main research content is briefly described as follows:(1)Four common types of memristor models are described in detail,and the corresponding Spice models are respectively built up,based upon their mathematical expressions.Meanwhile,the electrical properties of these four memristor models are discussed through a series of circuit simulations with the relevant contrastive analysis.Next,the specific memristance variable rules under the fixed-amplitude external stimulus are given.Furthermore,the scope of applications for different memristor models are clearly summarized,which provides the theoretical and experimental support for the subsequent memristor-based applications.(2)Based on the memristor nonlinear ion drift model,the corresponding charge-controlled and flux-controlled memristor models are presented respectively.Meanwhile,when the different external stimuluses are applied to the memristive element,the relevant time estimation method for the memristance variation is put forward,based on the comprehensive formula derivation.Furthermore,the composite behaviors of the multiple memristor circuits(i.e.,the memristors connected in series,parallel,hybrid series-parallel,and crossbar array configurations)are investigated.(3)According to the basic properties of the memristor based circuit,a hybrid complementary metal-oxide-semiconductor(CMOS)/memristor synaptic circuit is proposed.According to the mathematical derivation and circuit analysis,the synaptic weight can be expressed in hardware.As for the proposed synaptic circuit,the issue of the lack of completeness in weight sign can be solved effectively,and the linearity in weight processing and weight programming can be easily achieved.Furthermore,the corresponding neuron circuit synthesized with multiple memristive synaptic circuits and an activation unit is further designed,which can be utilized to constitute a compact multilayer neural network with fully-connected configuration(including the description of a hardware-friendly chip-in-the-loop learning method during the network training phase).For the verification purpose,the presented neural network is applied for the realization of single image super-resolution reconstruction.(4)For the classical pulse coupled neural network(PCNN),a novel memristor crossbar array with its necessary peripheral circuits is proposed,which is able to simulate the adaptive-variable linking coefficient in PCNN.Meanwhile,a flexible mapping function is designed,which makes the presented memristive PCNN achieve a good network performance in different applications.Finally,the validity and effectiveness of the entire scheme is verified by a case study on multi-focus image fusion,which offers potential benefits in addressing the inherent parameter estimation issue emerging in neuromorphic systems.