Unifying model explainability and robustness via reasoning labels

Explainability in deep learning has emerged as an important topic in recent years, with several works exploring various notions and mechanisms of explainability for deep neural networks (DNNs). In this paper, we draw upon the insight that in many situations model explainability is a means to assess another related yet distinct criterion model robustness. In order to render the link between explainability and robustness more explicit, we propose to use human-understandable reasoning labels during the training process of DNNs. The reasoning labels are jointly learned with the traditional classification labels. This joint training enables the model to predict a set of reasoning labels with every predicted class label. Then, we tie model explainability and robustness by introducing a notion of prediction consistency, whereby the model predictions are accepted—or considered robust—only when the predicted class and the predicted reasoning labels follow a certain pre-specified mapping. We show that by adopting such a framework, one can improve the classification accuracy of the state-of-the-art models (on consistent samples). We further show that using this notion of consistency makes the model more robust to adversarial perturbations.