Dottorato in INGEGNERIA INFORMATICA

Titolo della tesi: Deep Learning Across Domains: From Forecasting Environmental Signals and Classifying Medical Data to Explainable AI Systems

The rapid development of Artificial Intelligence and Deep Learning techniques has profoundly transformed a wide range of application domains, from medical diagnostics to environmental monitoring. However, several challenges remain in adapting advanced models to domains where data are scarce, heterogeneous, or require strong interpretability. This dissertation presents a series of research contributions addressing these issues across three main axes. First, in the medical domain, where I mainly investigated the intersection of Raman spectroscopy and AI for cancer diagnostics, providing a comprehensive survey and critical analysis of how machine learning can support cancer grading. This work highlights both the potential and the limitations of applying AI methods in sensitive applications, where ethical considerations are essential. Second, I explored time series forecasting in environmental contexts, with a focus on air quality and sea temperature signals. Leveraging deep learning and signal transformation techniques, I developed methods that improve predictive accuracy and robustness in environmental monitoring tasks, contributing to the development of reliable tools for sustainability and public health. Finally, my research turned towards explainable artificial intelligence, evaluating the role of explanations as both a tool for knowledge distillation and a possible vulnerability. I investigated how explainability can support model compression while also examining the risks of explanation manipulation, thus contributing to the growing discussion on trustworthiness and reliability of AI systems. Together, these contributions illustrate multiple trajectories from domain-specific applications toward methodological advances in generalizability and performance. Beyond the immediate findings, this dissertation underscores the adaptability and breadth of deep learning approaches across diverse domains, thereby contributing to the continued advancement of artificial intelligence.