Dual Input Capsule Network for Image Recognition

Abstract

Performance evaluation is a critical part of deep learning (DL) that requires careful conduct to enhance confidence and reliability. Several metrics exist to evaluate DL models, however, choosing one for a given model is not trivial, since it is not a one-fit-all solution. Practically, accuracy is the most popularly used evaluation metric for capsule networks (CapsNets). This is problematic for sensitive applications (such as those in health), since accuracy is overly optimistic in the presence of class imbalance (a common problem in health), and does not permit the exact reporting of a model’s risk of bias and potential usefulness. Besides, decisions in health are based on the estimated risks, bias, and the probability that a condition is present. It is, therefore, necessary to complement a model’s accuracy with a measure of its ability to: fail safely, determine regions of interest, measure the effectiveness of the DL algorithm, properly reconstruct the input images, and effectively extract features. The feasibility of this approach is experimentally shown in this paper with the implementation of two CapsNet models. The methods suggested in this paper do not only measure performance, but also contribute largely to model interpretability, understandability, and reliability