The paper introduces two efficient approximations for standard convolutional neural networks: Binary-Weight-Networks and XNOR-Networks. Binary-Weight-Networks approximate filters with binary values, reducing memory usage by 32 times. XNOR-Networks further approximate both filters and inputs to convolutional layers with binary values, enabling convolutions to be performed using primarily binary operations. This results in a 58 times faster convolutional operation and a 32 times reduction in memory usage. The authors evaluate their approach on the ImageNet classification task, showing that a Binary-Weight-Network version of AlexNet achieves the same classification accuracy as the full-precision version while being significantly more efficient. They also compare their method with recent network binarization methods, BinaryConnect and BinaryNets, and outperform them by more than 16% in top-1 accuracy. The code for their method is available at http://allenai.org/plato/xnornet.The paper introduces two efficient approximations for standard convolutional neural networks: Binary-Weight-Networks and XNOR-Networks. Binary-Weight-Networks approximate filters with binary values, reducing memory usage by 32 times. XNOR-Networks further approximate both filters and inputs to convolutional layers with binary values, enabling convolutions to be performed using primarily binary operations. This results in a 58 times faster convolutional operation and a 32 times reduction in memory usage. The authors evaluate their approach on the ImageNet classification task, showing that a Binary-Weight-Network version of AlexNet achieves the same classification accuracy as the full-precision version while being significantly more efficient. They also compare their method with recent network binarization methods, BinaryConnect and BinaryNets, and outperform them by more than 16% in top-1 accuracy. The code for their method is available at http://allenai.org/plato/xnornet.