Understanding and Exploiting the Latent Space to improve Machine Learning models eXplainability

In recent years, Artificial Intelligence (AI) and Machine Learning (ML) systems have dramatically increased their capabilities, achieving human-like or even humansuperior performance in specific tasks. This increased performance has gone hand in hand with an increase in the complexity of AI and ML models, compromising their transparency and trustworthiness and making them inscrutable black boxes for decision making. Explainable AI (XAI) is a field that seeks to make the decisions suggested by ML models more transparent to human users, by providing different types of explanations. This thesis explores the possibility of using a reduced feature space called “latent space”, produced by a particular kind of ML models, as a means for the explanation process. First, we study the possibility of navigating the latent space as a form of interactive explanation to better understand the rationale behind the model’s predictions. Second, we propose an interpretable-by-design approach to make the explanation process completely transparent to the user. Third, we exploit mathematical properties of the latent space of certain ML models (similarity and linearity) to produce explanations that are shown more plausible and accurate than those of existing competitors in the state of the art. In order to validate our approach, we perform extensive benchmarking on different datasets, with respect to both existing metrics and new ones introduced in our work to highlight new XAI problems, beyond current literature.