27TH EUROPEAN CONFERENCE ON ARTIFICIAL INTELLIGENCE

Robust Learning Methods for Imprecise Data and Cautious Inference

Abstract: The representation, quantification and management of uncertainty is a central problem in AI, and particularly so in machine learning (ML). Among different forms of uncertainty, imprecision, that is, the problem of dealing with imperfect and incomplete data, has recently attracted interest in the research community, for its implications on ML practice. The talk will focus on the problem of dealing with imprecision in ML from two different perspectives. On the one hand, imprecision affecting the input data to an ML pipeline, leading to the problem of learning from imprecise data. On the other hand, imprecision used as a way to implement uncertainty quantification for ML methods, so-called cautious inference methods. Within the context of learning from imprecise data, the talk will focus on how to formally represent and study learning from imprecise data problems, including simple yet effective algorithms that are able to learn despite having access only to partial information. Within the context of cautious inference, the talk will focus on an investigation of the conformal prediction framework, a theoretically-motivated cautious inference framework inspired by algorithmic information theory, and on how cautious inference could be used to enable safer, human-centred decision-making in human-AI interaction.

Bio: Andrea Campagner is a researcher at IRCCS Ospedale Galeazzi Sant'Ambrogio and a lecturer at the University of Milano-Bicocca. His research focuses on uncertainty management, machine learning, human-AI interaction, and medical informatics. He won Best Paper awards at several conferences, and the ACM SIGCHI Gary Marsden and IJAR Early Career Researcher awards. He is an Associate Editor of the International Journal of Medical Informatics and of Soft Computing, and he served as PC Chair of IJCRS-2023.