Real-time Image Recognition System using RRAM based Neural Network

Abstract

In this paper, the convolution operation repeatedly performed in Convolution Neural Network (CNN) is implemented in hardware using Resistive Random Access Memory (RRAM) and neuron circuits. Vector Matrix Multiplication (VMM) operation is performed in parallel using RRAM, and operation result of VMM is converted into pulse outputs using stochastic neuron circuits. To reduce the computation time of stochastic neuron circuits and image processing time, modified the ‘LeNet-5’ neural network model with input data pooling and increase of convolution operation strides. parameters of neural network are learned based on MNIST (Modified National Institute of Standards and Technology) data set. Weights of neural network are replaced with conductance of RRAM through transfer learning, and real-time image recognition of MNIST images is confirmed by combining with camera module. This neuron circuit has been designed with 0.18μm General Logic process by MagnaChip, and simulation verification performed by CADENCE Virtuoso circuit simulation tool and MATLAB.