Sebastian Palacio Bustamante

• The rising demand for machine learning (ML) models has become a growing concern for stakeholders who depend on automatic decisions. In today's world, black-box solutions (in particular deep neural networks) are being continuously implemented for more and more high-stake scenarios like medical diagnosis or autonomous vehicles. Unfortunately, when these opaque models make predictions that do not align with our expectations, finding a valid justification is simply not possible. Explainable Artificial Intelligence (XAI) has emerged in response to our need for finding reasons that justify what a machine sees, but we don't. However, contributions in this field are mostly centered around local structures such as individual neurons or single input samples. Global characteristics that govern the behavior of a model are still poorly understood or have not been explored yet. An aggravating factor is the lack of a standard terminology to contextualize and compare contributions in this field. Such lack of consensus is depriving the ML community from ultimately moving away from black-boxes, and start creating systematic methods to design models that are interpretable by design. So, what are the global patterns that govern the behavior of modern neural networks, and what can we do to make these models more interpretable from the start? This thesis delves into both issues, unveiling patterns about existing models, and establishing strategies that lead to more interpretable architectures. These include biases coming from imbalanced datasets, quantification of model capacity, and robustness against adversarial attacks. When looking for new models that are interpretable by design, this work proposes a strategy to add more structure to neural networks, based on auxiliary tasks that are semantically related to the main objective. This strategy is the result of applying a novel theoretical framework proposed as part of this work. The XAI framework is meant to contextualize and compare contributions in XAI by providing actionable definitions for terms like "explanation" and "interpretation." Altogether, these contributions address dire demands for understanding more about the global behavior of modern deep neural networks. More importantly, they can be used as a blueprint for designing novel, and more interpretable architectures. By tackling issues from the present and the future of XAI, results from this work are a firm step towards more interpretable models for computer vision.