XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks

XNOR-Net introduces two efficient approximations to standard convolutional neural networks aimed at drastically reducing memory and computation. In Binary-Weight-Networks, the convolutional filters are approximated with binary values, which results in roughly 32x memory savings. In XNOR-Networks, both the filters and the inputs to the convolutional layers are binary, so convolutions are computed using primarily binary operations, yielding about 58x faster convolution in terms of high-precision operations along with the 32x memory savings.

These efficiencies make it possible to run state-of-the-art networks on CPUs rather than GPUs in real time while remaining simple and accurate. Evaluated on ImageNet classification, a Binary-Weight-Network version of AlexNet achieves the same accuracy as the full-precision AlexNet, and the approach outperforms prior binarization methods BinaryConnect and BinaryNet by large margins - more than 16% in top-1 accuracy - establishing binary networks as viable for challenging visual tasks.