Weight Compression-Friendly Binarized Neural Network

Abstract

The resources of edge devices in AIoT systems are usually constrained with size and power. The computational complexity of neural network models in these edge devices has become a major concern. The most compact form of deep neural networks is binarized neural network (BNN), which adopts binary weights and exclusive NOR (XNOR) operations as binary convolution. In this paper, we propose weight compression-friendly BNN to save hardware resources by reducing memory space. The proposed technique does not binarize weights just according to the signs of weights, but fully considers compression efficiency in the training of the BNN model. The experiments are performed by using the binary version of a 6-layer convolutional neural network (CNN) and MNIST case. The results show that the proposed technique can achieve more than 25% reduction in memory space with accuracy loss of 1 %, or more than 35% memory reduction with about 2.5% accuracy drop for MNIST classification. The weight compression method does not destroy the regular structure of neural networks, so the proposed technique is very fit for processor-based BNN hardware accelerators.