Post-Hoc, Contrastive, Explainable Artificial Intelligence for Time Series and Image Data

Recent advances in science and technology have fostered the extensive deployment of artificial intelligence (AI) systems for prediction and decision making in our daily lives. However, there are growing concerns about the ability of these intelligent systems to explain what they do. This problem is especially evident in opaque black-box systems, such as deep neural networks (DNNs), which boast state-of-the-art performance in a variety of domains. Driven by the widespread application of these systems in high-stakes scenarios, recent government regulations are requiring AI systems to explain their output (e.g., the General Data Protection Regulation [GDPR] in the European Union [EU]) and ensure fairness, equality, and transparency in the scientific community. As such, there is an urgent demand for solutions in eXplainable AI (XAI). Consequently, XAI has become one of the most prominent research areas in recent years, with numerous conferences, workshops, industry research programs, and government policies dedicated to this topic. While progress has been made in explaining the predictions of black-box models for tabular and, to some extent, image data, solutions for time series data are less common. To date, most work in XAI has been on post-hoc factual explanations using feature importance, saliency mapping and examples (i.e., nearest neighbors of query instances). While factual explanations focus on why a certain prediction was made, contrastive counterfactual explanations look to how that prediction could have been different. This thesis explores the promise of contrastive explanations for black-box models operating with time series and image data. A core focus of the contribution is on post-hoc counterfactual explanations and the role of user evaluation. There is a growing consensus that counterfactual explanations are causally informative, psychologically effective, and legally compliant with respect to GDPR. Arguably, counterfactuals provide more robust and informative explanations than feature-importance methods. The results of this thesis demonstrate that (1) counterfactual explanations can be effective for time series and image data, (2) the properties of good explanations do not readily translate from tabular data to images and time series data, and (3) user-centric methodologies and evaluations are central to supporting continued progress in XAI.