Boosting the Robustness of Capsule Networks with Diverse Ensemble

The recent proposed Capsule networks encounter performance degeneration for the distorted images, which challenges the robustness of the models. To address this issue, we propose to enhance the robustness of capsule by constructing ensembles of the individual networks. By exploring the interaction among each individual network, we further propose to improve the robustness for the ensemble capsules by promoting the diversity during the training process. Extensive results on the various noises demonstrate that our algorithm can improve the robustness for the vanilla capsule networks in terms of the Gaussian as well as the adversarial perturbations.