Towards the Characterization of Representations Learned via Capsule-based Network Architectures