Abstract—Capsule Networks (CapsNets) have been well known for its part-whole relational property, whilst with heavy computation of the capsule routing. The classic Expectation-Maximization (EM) capsule routing first compute the vote matrix by multiplying part-whole pose matrix and learnable weight matrix, and secondly fed vote matrices and activations into EM algorithm for clustering. The heavy computation comes from a large-scale computations of vote matrices and large-scale data EM cluster. To address the challenge of lightweight design of CapsNets, different from the previous EM capsule routing that obeys the first-vote-then-cluster rule, we implement a novel first-cluterthen-vote mechanism. To this end, in this paper, we develop a capsule-type normalization routing algorithm as illustrated in Fig. 1 (b). Specifically, we first normalize the part-level capsules along the type dimension with the aim of transforming all types of capsules into a uniform distribution. Secondly, all part-level capsules and their mixed capsules are voted to the whole-level capsules via a multiplication with a single transformable matrix. In such way, the cluster computation and matrix multiplication computation both get reduced with a large margin. Our capsuletype normalization routing builds a deep CapsNets, which are proved to be promising on multiple datasets, including MNIST, SVHN, SmallNORB, CIFAR-10/100. Notably, our CapsNets can be implemented on the large-scale ImageNet-1K dataset and beats ResNets, which is quite difficult for the previous CapsNets versions. 如果你想下载本文相关的详细 PDF 文件，可以点击下面的链接进行下载：