Analysis of binary weighted network and XNOR-net binarized convolutional deep neural networks

Convolutional deep neural networks (CNN) are mostly used in image recognition and image detection processes and show impressive performances. The dimensions of CNN models are expanding day by day. In this way, while relatively increasing the successes, the high number of parameters, memory and transaction costs also increase. Due to this high memory and processing cost, problems occur in the limited hardware such as application of embedded systems, mobile devices, and IoT (Internet of Things). Various solutions are developed for these problems. One of the methods introduced to reduce the high memory and processing costs of the CNNs is to binarize the network. GPU (Graphic Processing Unit) is required for most of the CNNs to work effectively. Only CPU (Central Processing Unit) is not enough. Binarized artificial deep neural networks can be solved for GPU dependency and make possible to operate the CNNs in mid - low performance CPUs. In this study, BWN (Binary Weighted Network) and XNOR-NET binarized neural network structures are examined. LeNet-5, AlexNet, VGG models and the models which specialized for this study were tested using CIFAR-10, CIFAR-100, MNIST and SVHN datasets. Comparisons were made with networks with classical decimal parameters based on criteria such as accuracy, processing speed, memory cost and energy consumption. CNNs work approximate 58 times faster with XNOR-NET. With both XNOR-NET and BWN, it is provided with 32 times memory saving compared to single precision floating point format and 64 times memory saving compared double precision floating point format.