• ai
• analysis
• coverage
• decision
• deep
• dnn
• explainability
• learning
• machine
• network
• neural
• robust
• safe
• tree
• xai

Drawing inspiration from coverage testing in software engineering, a number of coverage criteria for Deep Neural Networks (DNNs) have emerged in recent years.
These criteria serve as metrics for conducting heuristic state searches within the domain of DNNs to identify potential unsafe behaviors.
Their purpose is to shed light on the reliability of the testing phase accuracy by quantifying the extent to which it can be trusted.
Over the past few years, there has been significant debate surrounding the efficacy of structural coverage metrics in enhancing DNNs, as there is a lack of evidence supporting their ability to enhance robustness.
This thesis delves into the exploration of whether structural coverage metrics exhibit greater effectiveness when employed in the context of more predictable machine learning models like Deep Neural Decision Trees (DNDTs) since the inherent interpretability of DNDTs enables a more convenient examination and explanation of the model's internal workings during the testing phase.
In particular, we present an architecture for using DNDTs in Computer Vision scenarios and we propose a new coverage metric that can be used over this architecture.