A brief survey on Capsule Network
Ruiyang Shi, Lingfeng Niu · 2020
Capsule networks(CapsNets) are new kinds of network representation in deep learning. They are proposed to overcome the shortcomings of convolutional neural networks(CNNs). CNNs do not consider the important spatial level correlation between simple and complex objects. Besides, the pooling operations lose too much spatial information. In contrast, CapsNets use a new architecture that mimics the human visual system to obtain equivariance, instead of the original translational invariance, so that they can use less data to get more extensive generalization in different perspectives. However, as a new research field, the lack of knowledge and the working principle of capsules are still obstacles for researchers to make big progress. In this paper, we review the related works of CapsNets including the structure design, routing mechanisms and applications. Finally, we discuss and summarize the advantages and disadvantages of CapsNets at present, and look forward to the future research directions.