Memristor-based Circuit Design Of Competitive Neural Network And Its Application On Pattern Recognition

Competitive neural network(CNN)is a significant class of Artificial neural network,which is broadly applied in pattern recognition for its unsupervised and self-organizing learning characteristics.The new nano-level component memristor can effectively simulate neural synapses.In this thesis,we try to design pure circuits based on memristor to realize CNN and apply them to pattern recognition.Firstly,a Memristive Competitive neural network(MCNN)circuit based on Hebbian learning rule is designed,which consists of forward calculation part and weight updating part.The forward calculation part adopts the winner-take-all rule to achieve competitive calculation,and the weight updating part adopts the Hebbian learning rule to achieve unsupervised weight adjustment.The function of the circuit is verified by the simulation experiment of image classification.Then,on the basis of this,a MCNN based on Spike Timing Dependent Plasticity(STDP)learning rule which is closer to Biological neural network is designed.The circuit consists of memristor cross array module,changed Leaky-integrate-and-fire module,post-spike generating module and signal switch module.The functions of four modules are verified through simulation experiments,and the functions of the overall circuit were also verified through simulation experiments of image classification.The performance of the circuit is compared with the MCNN circuit based on Hebbian learning rule.Finally,a bilinear interpolation circuit based on memristor cross array is designed,which is combined with the MCNN circuit based on STDP learning rule to realize the recognition of handwritten characters.It is concluded that the designed circuit is suitable for complex pattern classification tasks with small number of categories.The main achievements of this thesis include the design of a MCNN circuit based on the Hebbian learning rule,a MCNN circuit based on the STDP learning rule,a bilinear interpolation circuit based on the memristor cross array and the realization of handwritten character recognition by combining linear interpolation circuit and MCNN circuit.All the designed circuits are implemented in PSPICE circuit simulation software.Compared with the existing works,the running process of the designed MCNN circuit does not require the participation of Central Processing Unit,Field-Programmable Gate Array,microcontroller and other control units.The hardware consumption is smaller.It provides a new method for the hardware implementation of CNN,and provides a new idea for the implementation of computing-in-memory and parallel computing architecture.