Adaptive capsule network

A capsule is a group of neurons whose outputs represent different properties of the same entity. Typically, the capsule is produced by applying convolution layers called the primary capsule layer to group scalar neurons. However, randomly grouping the scalar neurons into capsule vectors can cause two problems: (i) The capsule vectors are difficult to obtain a better representation of the entities. (ii) The capsule vectors generated by the primary capsule layer lack spatial information and cannot effectively model the underlying spatial relationship among entities. In this paper, we present a flexible and efficient capsule network architecture called Adaptive Capsule CapsuleNet (AC-CapsNet). We replace the primary capsule layer of CapsNet with the adaptive capsule (AC) layer. In the AC-CapsNet, the adaptive capsule vector combines capsule vector and adaptive value generated by the AC layer. The adaptive value preserves spatial information of each capsule vector and local relationship among the scalar neurons contained in each capsule vector. Therefore, the adaptive capsule vector can not only dynamically adjust their state values according to the content information of scalar neurons inside capsule vector, but also model the spatial relationship between capsule vectors for low-level clusters. Extensive experiments in some public datasets such as CIFAR-100, CIFAR-10, SmallNORB, and SVHN show that the AC-CapsNet outperforms other variants of CapsNets with respect to classification accuracy and robustness to affine transformations and white-box adversarial attacks.