Abstract

In this thesis, we explore the explainable robustness of neural networks for visual classification. We study an essential question for making neural networks deployable in real-world applications: “how to make neural networks explainably robust?” We start by making neural networks explainable. It begins with enabling black-box neural networks to justify their reasoning by leveraging attributes, i.e., visually discriminative properties of objects, and perturbations, to provide counterfactual explanations. The two chapters that follow focus on enhancing the robustness of neural networks against natural and adversarial perturbations. We do so by integrating perturbations in the network architecture and provide a rationale behind the modification of the network for enhancing its robustness by training the standard network with similarly transformed images. The last chapter utilizes attributes to improve robustness against perturbations and provides explanations as a byproduct.