Trustworhy AI through uncertainty quantification

Machine Learning (ML) is nowadays becoming widespread in many applicative fields. In many cases, however, ensuring a certain degree of reliability in the system's prediction is paramount, especially in life-critical applications such as autonomous driving and medical diagnosis. One possible answer to this requirement is to quantify the \textit{level of uncertainty} associated with the model's predictions: indeed, in principle, such quantified value can be used for understanding when to and when not to \textit{trust} the AI system, allowing for more informed choices. Thus, Uncertainty Quantification (UQ) is gradually becoming critical to ensure the trustworthiness of many modern AI-based systems. Most state-of-the-art UQ techniques are general-purpose strategies (i.e., not depending upon the task at hand) that require a specific adaptation at the training stage; as a result, UQ comes at the price of developing an ``ad hoc'' uncertainty-aware training that fulfills the conditions for constructing an uncertainty measure. Nevertheless, distinct applications have different budgets: on the one hand, such a condition is not applicable within industrial applications for which the cost of re-training is too high; on the other, this solution is sub-optimal whenever it is possible to enforce the notion of uncertainty by exploiting task-specific characteristics (which requires a higher budget). Since the budget may considerably depend upon the security requirements, this thesis tries to see UQ in a new light, i.e., not as a monolithic ad hoc general-purpose answer but as a spectrum that ranges from the most general to the most application-specific solution. Accordingly, this thesis aims to explore the untried regions of the uncertainty spectrum with particular attention to the potential needs of concrete applications. We start this exploration with post hoc UQ techniques, i.e., which act on an already trained neural network. We developed a theoretically founded strategy by using a sigma-scaled uncertainty measure derived from MC-dropout envisaged on already trained neural networks (namely, Dropout Injection). Then, we explore application-driven strategies for UQ on density regressors, exploiting the non-negative nature of the outputs in this domain by fitting a Rectified Gaussian distribution Before the ReLU Estimates (BLUES Bayesian Inference). Finally, we conduct a comparative study on the trustworthiness of such techniques to shed light on their feasibility in adversarial domains.