Circuit Design Of Memristor Synapse Based In-situ Training Neural Network

With the development of machine learning theory,the algorithm of artificial neural networks has been widely used in many application fields such as image recognition,natural language processing,pattern classification,decision control,and achieved an excellent performance.However,use traditional digital computers based on von Neumann architecture to implement artificial neural networks will generate large power consumption that restricts the application and development of artificial neural networks from the hardware aspect.A new circuit element called memristor with memory characteristics has broad research prospects in many areas like memory device,chaotic circuit design,and cell synapse function simulation.Using memristors can construct the non-von Neumann computing architecture and realizing in-memory computing,which significantly improves the energy efficiency ratio of computing devices.The uses of memristor provide a new solution to solving the hardware limitation of artificial neural networks,and also offer a new method to realize bionic brain-like computing.This thesis demonstrates the design flow of memristor-based neural network circuit system in four aspects: the selection of memristor model,the improvement of memristorbased synapse circuit,the design of the in-situ training circuit and the circuit implement of the completed neural network.Firstly,the basic theory and commonly used models of memristor are introduced as the foundation of memristor-based neural network design.How the memristor model’s change characteristics affect to neural network’s performance is analyzed.Then,considering the representation rang of synaptic weight and the synaptic weight reading/writing accuracy,this thesis proposes an improved memristor synapse circuit.The proposed memristor synapse circuit can meet the requirements of large-scale neural network circuits design.A novel in-situ training circuit design using the feedback control strategy is also proposed based on the proposed synapse circuit.The proposed insitu training circuit needs not to encode the write pulse according to memristor’s change characteristics,and it directly controls the learning process depend on the real-time state of the memristor.Finally,this thesis designs the specific functional circuit block to realize the network’s various operations like neuron activation and weight update according to the bidirectional associative memory neural network’s algorithm.Using these designed functional blocks can build the completed memristor-based neural network circuit system with learning and using function.The constructed memristor-based neural network circuit system proves its high performance through the comparison test of software-based model and circuit-based model.Because the use of feedback control strategy can avoid the encoding error due to the memristor’s nonlinear characteristics,making the proposed in-situ training circuit is suitable for the realistic fit memristor model.The adjustment error caused by circuit’s delay parameters during learning process is discussed.How to reduce the total error is also given.Compared with other similar works,the memristor-based neural network circuit system proposed in this thesis has the advantages of fewer memristor quantities and all memristors’ synchronous weight change,making the neural network circuit system in this thesis has higher learning efficiency.