Balanced Circulant Binary Convolutional Networks

Abstract

Binary convolutional neural networks (BCNNs) are widely used to improve memory and computation efficiency of deep convolutional neural networks (DCNNs) for mobile and AI chips based applications. However, current BCNNs are not able to fully explore their corresponding full-precision models, causing a significant performance gap between them. In this paper, we propose balanced circulant binary convolutional networks (BCBCNs), towards optimized BCNNs, by balancing the distribution of feature maps while enhancing the orientation ability of kernels. In particular, we adjust the architecture by introducing more batch normalization (BN) layers and circulant convolutional layers in an end-to-end framework, which significantly improve the performance of BCNNs. This combination can be easily exploited into existing DCNNs such as LeNet and ResNet. Extensive experiments demonstrate the superior performance of the proposed BCBCNs over most state-of-the-art BCNNs.