OPENING THE BLACK BOX: AN ACTIVE APPROACH TO EXPLANATION METHODS IN MACHINE LEARNING
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
With the popularity of machine learning methods ever on the rise, it is more important than ever to create models that can be trusted. This dissertation presents an integrated active framework for explainable AI (XAI) that enhances the transparency of machine learning models, which are especially critical in high-stakes decision-making scenarios. The importance of model transparency is emphasized, as opaque decision-making processes in AI can lead to mistrust and ethical concerns. In this research, we include an in-depth exploration of the current landscape of explanation methods within the XAI field, paying particular attention to feature-importance-based methods, which provide insights into model decision-making by highlighting the relevance and impact of different features on model predictions. Through this exploration, we discuss key challenges and opportunities for improving explainability in machine learning.
In addition to the introduction of two novel feature-importance-based explanation methods, the core contributions of this dissertation are: (1) our investigation into the nature and solvability of the weakly supervised object localization (WSOL) problem and (2) an introduction of an integrated active framework for active XAI, wherein models are trained with an active awareness of their own explainability in relation to a chosen explainer. This approach ensures that models trained within this framework are not only accurate but also maximizes interpretability and transparency in its decision-making process (i.e., attain minimum risk with maximum explainability). This framework contributes to the ongoing efforts to enhance the trustworthiness and accountability of AI systems, paving the way for more responsible and ethical AI deployment in high-stakes domains.