A taxonomy of interpretation and explanation methods for capsule network architectures

Abstract

In critical domains such as healthcare, autonomous driving, environmental monitoring, and legalization and analysis of official documents, the necessity for interpretable machine learning models is of utmost importance. The capability to explicate the mechanisms by which a model arrives at its predictions is essential for fostering trust, ensuring transparency, and enabling informed decision-making. This imperative requires a transition from conventional “black box” models to intrinsically interpretable architectures. Capsule Neural Networks (CapsNets) represent a promising avenue in this regard. Unlike classical Convolutional Neural Networks (CNNs), which often lack transparency due to their feature abstraction (where deep layers capture complex representations their reliance on pooling operations, which lose hierarchical relationships between object parts) and black box characteristics (it is hard to trace how specific features influence predictions), CapsNets are designed to learn object-centric representations. This is achieved by explicitly modeling hierarchical relationships within the network. This provides a more intuitive understanding of the decision-making process. To analyze the model trained in such sensitive domains, there are two main tasks: model interpretation, which involves examining the inner workings of the model. Model explanation, which focuses on justifying model predictions. This paper studies interpretation and explanation methods designed for CapsNets trained across diverse problems. Additionally, we introduce a set of factors used to position these methods into four main categories, providing researchers with a terminology to characterize their works and position them with respect to existing efforts. Our study shows that, in the context of CapsNets, model interpretation has received more attention, while the explanation task has remained relatively unexplored. We expect that this study will serve as a valuable point of reference for the research community, positioning the capabilities of CapsNets at the forefront of modern research.