Pairwise Coupling of Convolutional Neural Networks for the Better Explainability of Classification Systems

Ensembling techniques are viewed as a promising machine learning tool to resolve issues arising from the monolithic nature of deep neural networks. In this paper, we consider pairwise coupling models built from neural networks, which is a special kind of ensemble. These models promise to provide much needed modularity in classification models employing deep networks. In order to be practical, pairwise coupling models have to have comparable memory and speed requirements to commonly used architectures. In this paper, we propose novel architectures that address this key problem of pairwise coupling models. We show that the classification accuracy of the resulting pairwise coupling models matches the original network, while exceeding the original network’s pairwise accuracy. The introduction of these pairwise models brings additional benefits. First, they allow for much less expensive uncertainty predictions. Secondly, their modularity allows for the fine-tuning of classification accuracy. Both of these benefits can be viewed as relating to the larger topic of improving the explainability as well as the modularity of deep neural networks.