Discovering the World Through Unsupervised Statistical Relational Learning

In the past decade, deep learning (DL) has significantly advanced machine learning (ML) and artificial intelligence (AI), powering systems like AlphaGo, GPT, and Alexa. These successes stem from both technological progress and theoretical strengths of DL, such as neural networks’ ability to learn complex functions hierarchically and benefit from unsupervised pre-training on large datasets. However, DL still falls short of human-level intelligence. While humans are able to learn from limited data, understand and explain observations, imagine fictitious scenarios, plan actions, solve problems, share/acquire knowledge through social interactions and adapt over time, DL and unsupervised learning (UL) solutions are still data hungry, they struggle to generalize to tasks different from the ones they are trained for, they are subject to robustness and lack of interpretability issues and they tend to catastrophically forget over time. There is a need for general AI solutions that can integrate DL/UL, good at recognizing patterns from noisy data, with symbolic approaches, able to manipulate symbols and to perform complex reasoning, to build trustworthy and more human-like AI systems.

Goal: The main goal of this project is to develop the foundations for unsupervised statistical relational learning (USRL), by integrating representation and structure learning in Statistical Relational AI (StarAI) models, and to apply the devised algorithmic solutions to applications ranging from rule discovery in general game playing, learning symbolic abstractions and causal reasoning in mathematical and cognitive science domains.