No abstract

A neural network fully implemented by memristive crossbar circuit is proposed and simulated, which can operate in parallel for the entire process. During the forward propagation, memristors in crossbar structure and a column of fixed-value resistors implement multiply-add operations. During the backward training, each memristor is tuned in conductance independently by training pulses, which implements weight/bias updating. A high recognition accuracy of 93.65% for hand-written numbers is achieved, which is comparable to that for software solution. The effects of the number of conductance states and the amplification of synaptic array circuit on the recognition accuracy are also investigated.