Deeper Capsule Network For Complex Data

Abstract

Capsule Network (CapsNet) addresses the problem of Convolutional Neural Network (CNN) by introducing dynamic routing between capsules. Our work further develops CapsNet in depth and performance. By introducing Convolutional Capsule Layer (Conv-Caps Layer), we deepen CapsNet, which greatly improves the performance. We also propose Capsule Pool (Caps-Pool), a new pooling operation, to reduce the number of parameters. This pooling operation preserves the full representation of features. In our experiment, our DeeperCaps model received so far the strongest result of CapsNet on the bigger dataset (Cifar-10). Our Caps-Pool reduces half of the parameters between layers while maintaining the performance. It also influences the output distribution to be intuitively more reasonable.