Efficient Memristive Binary Neural Networks With Spatial Separable Convolutions

The binary neural network (BNN) is an attractive model for edge computing, and memristor-based neuromorphic computing platform provides competitive end-side deployment scheme. However, current BNNs still have massive convolutional kernel parameters, which remain an obstacle to their deployment on end-side computing platform. In this paper, an efficient BNN called BSSC-Net is proposed, which applies binary spatial separable convolution (BSSC) to reduce the parameters of the binary convolutional kernels. Furthermore, a BNN training framework called KDSNG-BNN is designed to improve the performance of BSSC-Net, which consists of a full-precision model, a BSSCNet, and a discriminator with spectral normalization layers. In this framework, our proposed BSSC-Net can improve the ability to learn feature distributions by simultaneously performing distillation and adversarial training. Experimental results show that the proposed BSSC-Net and BSSC-Net-I based on ResNet-18 trained in KDSNG-BNN can achieve 93.36% and 93.32% accuracy on CIFAR-10, outperforming ReActNet by 0.81% and 0.77%, respectively, and the convolutional kernel parameters of the BSSC-Net and BSSC-Net-I are only 81% and 76.3% of ReActNet. Meanwhile, current memristor technology is more suitable for supporting the computation of low-bit parameters, while BNN quantizes weights and activations to 1 bit, which is ideal for deployment on memristors. Therefore, this paper presents a memristive implementation scheme by exploiting the advantages of memristors and BSSC-Net, intending to realize and promote end-side intelligent applications.