Learning compositional capsule networks

Abstract

Objects in the visual field are perceived to have an inherent structure that is seen in the way that they are constructed from their components. For example, a face requires its parts to be arranged in a certain spatial configuration. This property, of having such a structure, is termed as compositionality. For deep neural networks to preserve these structures of their inputs in their representations, the capsule network model was proposed. However, there is no empirical evidence to confirm if capsule networks do indeed learn compositional representations. Here, we propose a novel task for the experimental analysis of this property. This task, termed MeasureComp, tests the unsupervised learning of unannotated part-whole structures in a classification setting. Our results show that capsule networks that use dynamic routing are unable to learn pose-aware representations. In an effort to improve upon this, and as an initial direction towards compositional capsule models, we propose a novel compositional loss-function termed EntrLoss. Experimental results on MeasureComp show that the use of this loss function improves the compositionality of capsule networks. Further, we also present a simple capsule network model that uses our EntrLoss and outperforms several other recent capsule networks. The code for our paper is available at https://github.com/codesubmissionforpaper/entropy_regularised_capsule.