XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks This paper proposes two efficient approximations of standard convolutional neural networks (CNNs): Binary-Weight-Networks and XNOR-Networks. Binary-Weight-Networks use binary values for filters, reducing memory usage by 32×. XNOR-Networks use binary values for both filters and inputs, resulting in 58× faster convolutional operations and 32× memory savings. These networks enable real-time inference on CPUs without GPUs, and are accurate on challenging visual tasks. The authors evaluate their approach on ImageNet classification, achieving the same accuracy as full-precision AlexNet with Binary-Weight-Networks. They outperform existing binarization methods like BinaryConnect and BinaryNet by a large margin, with a 16.3% improvement in top-1 accuracy. The code is available at http://allenai.org/plato/xnornet. The paper introduces binary CNNs by binarizing weights and intermediate representations. Binary-Weight-Networks approximate real weights with binary values, reducing memory usage and computational cost. XNOR-Networks further binarize inputs, enabling efficient convolutional operations using XNOR and bitcounting. The authors demonstrate that their methods achieve comparable accuracy to full-precision networks while being significantly more efficient. They also show that their methods outperform existing binarization techniques on ImageNet. The paper compares their methods with related work in deep learning, including shallow networks, network compression, and parameter quantization. They find that their methods are more efficient and accurate than existing approaches. The authors also perform ablation studies to evaluate the effectiveness of their methods, finding that scaling factors and block structures are crucial for performance. The paper concludes that their binary approximations of CNNs are simple, efficient, and accurate. They propose XNOR-Net, an architecture that uses mostly bitwise operations to approximate convolutions, providing 58× speedup and enabling real-time inference on CPUs. The authors also show that their methods outperform existing techniques on ImageNet, achieving high accuracy with reduced computational and memory requirements.XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks This paper proposes two efficient approximations of standard convolutional neural networks (CNNs): Binary-Weight-Networks and XNOR-Networks. Binary-Weight-Networks use binary values for filters, reducing memory usage by 32×. XNOR-Networks use binary values for both filters and inputs, resulting in 58× faster convolutional operations and 32× memory savings. These networks enable real-time inference on CPUs without GPUs, and are accurate on challenging visual tasks. The authors evaluate their approach on ImageNet classification, achieving the same accuracy as full-precision AlexNet with Binary-Weight-Networks. They outperform existing binarization methods like BinaryConnect and BinaryNet by a large margin, with a 16.3% improvement in top-1 accuracy. The code is available at http://allenai.org/plato/xnornet. The paper introduces binary CNNs by binarizing weights and intermediate representations. Binary-Weight-Networks approximate real weights with binary values, reducing memory usage and computational cost. XNOR-Networks further binarize inputs, enabling efficient convolutional operations using XNOR and bitcounting. The authors demonstrate that their methods achieve comparable accuracy to full-precision networks while being significantly more efficient. They also show that their methods outperform existing binarization techniques on ImageNet. The paper compares their methods with related work in deep learning, including shallow networks, network compression, and parameter quantization. They find that their methods are more efficient and accurate than existing approaches. The authors also perform ablation studies to evaluate the effectiveness of their methods, finding that scaling factors and block structures are crucial for performance. The paper concludes that their binary approximations of CNNs are simple, efficient, and accurate. They propose XNOR-Net, an architecture that uses mostly bitwise operations to approximate convolutions, providing 58× speedup and enabling real-time inference on CPUs. The authors also show that their methods outperform existing techniques on ImageNet, achieving high accuracy with reduced computational and memory requirements.