Co-PI: Nicolò Cesa-Bianchi (Università degli Studi di Milano Statale)
Adaptability has been one of the primary objectives of artificial intelligence since the beginning and is now widely necessary, not only in the virtual world but also in the physical world. In particular, adaptability is related to an entity’s ability to interact with the environment, to perceive the context and related information that changes over time, and to act promptly. In particular, machine learning has emerged so far as the main enabling technology for designing adaptive artificial systems.
Despite its ubiquitous adoption, there are serious gaps in our theoretical understanding of how learning systems can provide assurances and how they can be effectively combined with other AI paradigms when designing adaptive artificial systems. Challenges to be addressed include: the foundations of learning theory for adaptability, understanding non-convex optimization for machine learning models (taking into account the interaction between the statistical and algorithmic aspects of the training process), the study of the characteristics of the specific context that can be used to provide better guarantees of adaptation.
