Stefano Teso
Senior Assistant Professor at CIMeC and DISI, University of Trento
There is an increasing need of Artificial Intelligence (AI) and Machine Learning (ML) models that can reliably output predictions matching our expectations. Models learned from data should comply with specifications of desirable behavior supplied or elicited from humans and avoid overconfidence, i.e., being aware of their own ignorance. Much of current attention focuses on pure deep learning models, such as large language models, which are however not designed to follow specifications, and provide no guarantees of doing so, not even when fine-tuned to follow human instructions. This project aims to leverage probabilistic circuits to guarantee reliable and efficient probabilistic reasoning in the wild, that is, even outside of the training distribution and of lab conditions in general. Stefano Teso (University of Trento) and Antonio Vergari (University of Edinburgh) develop their ongoing collaboration ideas and techniques to encourage deep learning models, including large language models, to behave compatibly with specifications, and tools for evaluating the reliability of models that satisfy this criterion.
Keywords: Trustworthy Machine Learning, Probabilistic Reasoning, Neuro-symbolic Integration
Scientific area: Artificial Intelligence
Bio: Stefano Teso is a senior assistant professor at the University of Trento. His research revolves around trustworthy AI, and especially ensuring AIs are aligned with the needs and preferences of their users. He has worked on offline and interactive machine learning, explainable AI, integrating learning and reasoning, learning symbolic constraints from data, and preference elicitation. His work covers both conceptualization and implementation of practical solutions, as well as theoretical foundations. Stefano obtained his PhD from the University of Trento in 2013 with a thesis about applications of statistical relational learning in proteomics. He has published several research papers at top AI and machine learning venues, and serves as reviewer, meta-reviewer and area chair for many AI and ML conferences and journals.
Visiting period: 10/04/2024 – 30/05/2024 at APRIL, School of Informatics, University of Edinburgh