| With the deepening of research on artificial intelligence,the scale and computational complexity of neural networks have gradually increased as their performance has gradually improved.At the same time,artificial intelligence is gradually popularizing,which makes research on the marginalization of artificial intelligence gradually become a focus.However,the performance of edge AI is limited by the inability to effectively improve device computing power.In this case,the demand for accelerating the operation of neural networks by breaking through the memory wall of traditional von Neumann architecture and using in-memory computing and other methods is increasing.Therefore,new devices such as RRAM(Resistive Random Access Memory)have entered people’s field of vision.The unique performance of these non-volatile memories provides new ideas and directions for building neural networks and accelerating neural networks.This thesis studies the method of using RRAM to construct convolutional neural networks.The main work is as follows: studying the structure and working principle of convolutional neural networks,analyzing the structure and electrical characteristics of RRAM devices,and then studying the method of using RRAM’s resistive characteristics to replace weights in convolutional neural networks to construct hardware-implemented convolutional neural networks.This thesis designs an architecture of a non-binary weight convolutional neural network implemented using RRAM based on FPGA,which includes methods for replacing weights with conductance and programming conductance,discretization encoding methods for data in neural networks,and compression of interlayer data,etc.In addition,this thesis also uses a line buffer structure to further accelerate convolution operations in neural networks.The proposed scheme is simulated in Vivado,and the prototype verification of the convolutional neural network based on RRAM is completed on FPGA.The result shows that the designed network can achieve 83.63% recognition accuracy of the test set of MNIST database. |
