Constructing Better Primary Capsules

As a promising variant of neural network, CapsNet works well with MNIST handwritten digit recognition, but inferior to CNN on complex datasets. Instead of proposing novel algorithm for dynamic routing which has been confirmed by previous literatures that the improvement of performance is limited, we focus our attention on how to construct effectively primary capsules. In this paper, for learning better representation we replace convolutional layers with pretraining layers, the extracting latent representation can be utilized to construct high effective primary capsules, which is capable of reaching the state-of-the-art performances on complex image datasets compared with other variants of CapsNet and reducing calculation simultaneously caused by fewer capsules. Our results fully demonstrate that CapsNet with better constructing primary capsules can work well with less memory demands on complex image datasets and present a new direction for realizing true potential of CapsNet.