Memristor-Based Design of Sparse Compact Convolutional Neural Network

Abstract

Memristor has been widely studied for hardware implementation of neural networks due to the advantages of nanometer size, low power consumption, fast switching speed and functional similarity to biological synapse. However, it is difficult to realize memristor-based deep neural networks for there exist a large number of network parameters in general structures such as LeNet, FCN, etc. To mitigate this problem, this paper aims to design a memristor-based sparse compact convolutional neural network (MSCCNN) to reduce the number of memristors. We firstly use an average pooling and $1\times 1$ convolutional layer to replace fully connected layers. Meanwhile, depthwise separation convolution is utilized to replace traditional convolution to further reduce the number of parameters. Furthermore, a network pruning method is adopted to remove the redundant memristor crossbars for depthwise separation convolutional layers. Therefore, a more compact network structure is obtained while the recognition accuracy remaining unchanged. Simulation results show that the designed model achieves superior accuracy rates while greatly reducing the scale of the hardware circuit. Compared with traditional designs of memristor-based CNN, our proposed model has smaller area and lower power consumption.