Davide Sferrazza

The growing integration of artificial intelligence (AI) across diverse sectors calls for models that are not only accurate but also trustworthy and interpretable. This need is especially pronounced when deploying AI directly on edge devices, which enables real-time data processing close to the source while leveraging large-scale pre-trained multi-modal models. This approach offers clear advantages in latency and privacy but also raises challenges in understanding the internal mechanisms of these models, ensuring transparency, avoiding reliance on spurious features, and supporting efficient, user-specific adaptation. This project aims to study and design multi-modal deep learning models that optimally balance accuracy, trustworthiness, and efficiency in edge intelligence settings. The project will address several use cases related to visual place recognition (VPR) and video understanding, maintaining an overarching edge-intelligence framework. It will explore the internal knowledge of VPR models using concept-based interpretability frameworks, aiming to identify what concepts models rely on, diagnose failure cases, understand cross-model similarities, quantify predictive uncertainty in frozen models without retraining, improving efficiency for edge deployment. We will also leverage concept bottlenecks and structured reasoning to model temporal dynamics and enhance transparency in video anomaly detection. The third year will be dedicated to adapting the designed models to distributed and federated settings, taking into consideration privacy and security.