Argumentative Explanation for Deep Learning: A Survey

Neural Networks (NNs) are often referred to as “black box” models, which has sparked increasing interest among researchers in understanding their internal workings. Computational argumentation (CA), a subfield of symbolic Artificial Intelligence (AI), has demonstrated notable advantages in the field of Explainable Deep Learning (XDL) recently. This is primarily due to the inherent compatibility between the dialectical nature of argumentation and the key elements of interpretation. In this paper, we provide an introduction to several Argumentation Frameworks (AFs) that serve as important connections between NNs and explanation. Furthermore, we present a comprehensive overview of existing argumentation-based explanatory approaches, covering both ante-hoc explanation and post-hoc explanation. Finally, current issues and future directions that can be explored in depth are discussed, from both academic research and business application perspectives.