Machine Learning Seminar: Dr. Emanuele Marconato, "Symbol Grounding in Neuro-Symbolic AI: A Gentle Introduction to Reasoning Shortcuts."."
- Sponsor
- Research Area of Artificial Intelligence
- Speaker
- Emanuele Marconato
- Contact
- Allison Mette
- agk@illinois.edu
- Views
- 4
- Originating Calendar
- Siebel School Speakers Calendar
Abstract: Neuro-symbolic (NeSy) AI combines neural perception with symbolic reasoning to produce predictions that respect prior knowledge, such as structural or safety constraints. By mapping low-level inputs to high-level concepts and reasoning over them, NeSy systems offer a promising path toward reliable and trustworthy AI. However, when concepts are not directly supervised, these models can develop Reasoning Shortcuts (RSs): they achieve high accuracy while grounding intermediate concepts incorrectly. Such shortcuts undermine interpretability, weaken out-of-distribution generalization, and ultimately compromise reliability.
In this seminar, I provide an accessible and structured overview of RSs in NeSy models. I explain their causes and consequences, review existing theoretical characterizations, and discuss current mitigation and awareness strategies, highlighting their strengths and limitations. Overall, I aim to offer a unifying perspective and lower the barrier to addressing RSs in practice. Finally, I will discuss potential open venues beyond current research.
Bio: Emanuele Marconato is an Assistant Professor at the University of Trento. He obtained his Ph.D. in 2025 as part of the first cycle of the Italian National PhD Program in Artificial Intelligence, being among its first graduates, under the supervision of Prof. Andrea Passerini and Stefano Teso. His research focuses primarily on Explainable AI, Neuro-Symbolic Integration, and Representation Learning, with particular emphasis on the identifiability of representations. He has published in top AI conferences such as NeurIPS, ICML, UAI, and AISTATS, and has contributed to journals including MLJ and Entropy. He also serves as an active reviewer for several AI conferences and journals.