Abstract:
Based on the characteristic that the resistance of double memristor structure can be adjusted linearly, a design scheme of SOFM neural network system based on double memristor is proposed. The scheme consists of preprocessing module, double memristor weight module, Euclidean distance operation module, neuron decision module and memristor weight update module. The double memristor weight module is composed of a double memristor unit and an amplification unit. The double memristor unit is composed of two memristors with the same structure and connected with the doping region. Compared with the single memristor structure, because the total resistance of the double memristor can remain unchanged, the linear adjustment of the memristor resistance can be realized. The Euclidean distance operation module is composed of subtraction circuit, square circuit and addition circuit. The Euclidean distance between the input voltage signal and the weight voltage signal can be calculated, so as to provide decision-making basis for the neuron decision-making module. By adjusting the voltage signal to control the weight voltage of the double memristor weight module, the training and testing of SOFM neural network can be completed. According to the scheme, a clustering experiment is carried out. The experimental results show that the adjustment of memristor resistance in the range of 0.7kΩ~1.1kΩ and weight voltage in the range of 0.55 V~0.85 V can be realized by the designed neural network system. The experimental results of clustering 10 training samples into 8 categories are realized, , and 8 test samples are clustered into 4 categories, which is consistent with the test results of SOFM neural network algorithm. The effectiveness of the designed circuit is verified.