XNOR-Net: ImageNet Classification Using Binary Convolutional Neural\n Networks

Abstract

We propose two efficient approximations to standard convolutional neural\nnetworks: Binary-Weight-Networks and XNOR-Networks. In Binary-Weight-Networks,\nthe filters are approximated with binary values resulting in 32x memory saving.\nIn XNOR-Networks, both the filters and the input to convolutional layers are\nbinary. XNOR-Networks approximate convolutions using primarily binary\noperations. This results in 58x faster convolutional operations and 32x memory\nsavings. XNOR-Nets offer the possibility of running state-of-the-art networks\non CPUs (rather than GPUs) in real-time. Our binary networks are simple,\naccurate, efficient, and work on challenging visual tasks. We evaluate our\napproach on the ImageNet classification task. The classification accuracy with\na Binary-Weight-Network version of AlexNet is only 2.9% less than the\nfull-precision AlexNet (in top-1 measure). We compare our method with recent\nnetwork binarization methods, BinaryConnect and BinaryNets, and outperform\nthese methods by large margins on ImageNet, more than 16% in top-1 accuracy.\n