From Hashing to CNNs: Training BinaryWeight Networks via Hashing.

Deep convolutional neural networks (CNNs) have shown appealing performance onvarious computer vision tasks in recent years. This motivates people to deployCNNs to realworld applications. However, most of state-of-art CNNs requirelarge memory and computational resources, which hinders the deployment onmobile devices. Recent studies show that low-bit weight representation canreduce much storage and memory demand, and also can achieve efficient networkinference. To achieve this goal, we propose a novel approach named BWNH totrain Binary Weight Networks via Hashing. In this paper, we first reveal thestrong connection between inner-product preserving hashing and binary weightnetworks, and show that training binary weight networks can be intrinsicallyregarded as a hashing problem. Based on this perspective, we propose analternating optimization method to learn the hash codes instead of directlylearning binary weights. Extensive experiments on CIFAR10, CIFAR100 andImageNet demonstrate that our proposed BWNH outperforms current state-of-art bya large margin.