Batch normalization (BN) is crucial for achieving state-of-the-art binary neural networks (BNNs). Unlike full-precision neural networks, BNNs restrict activations to discrete values {-1,1}, which requires a renewed understanding and research of the role and significance of the BN layers in BNNs. Many studies notice this phenomenon and try to explain it. Inspired by these studies, we introduce the self-attention mechanism into BN and propose a novel Attention-Based Batch Normalization (ABN) for Binary Neural Networks. Also, we present an ablation study of parameter trade-offs in ABN, as well as an experimental analysis of the effect of ABN on BNNs. Experimental analyses show that our ABN method helps to capture image features, provide additional activation-like functions, and increase the imbalance of the activation distribution, and these features help to improve the performance of BNNs. Furthermore, we conduct image classification experiments over the CIFAR10, CIFAR100, and TinyImageNet datasets using BinaryNet and ResNet-18 network structures. The experimental results demonstrate that our ABN consistently outperforms the baseline BN across various benchmark datasets and models in terms of image classification accuracy. In addition, ABN exhibits less variance on the CIFAR datasets, which suggests that ABN can improve the stability and reliability of models.
Keywords: Binary neural networks; batch normalizationa; convolutional neural networks; deep learning.