TP5: Frontiers of Machine Learning – Hard Sciences for Machine Learning

The foundational contribution of the Hard Sciences to an advanced modeling and understanding of various aspects of machine learning beyond the current empirical settings is an undisputedly crucial aspect in the future development of AI in the next decades. This is even more so while moving toward the quantum realm that requires a completely new theoretical bootstrap.

The Transversal Project 5 will address the following foundational research question: How, why and to what extent do modern machine-learning models work? Transversal Project 5 will tackle such challenge with a strongly interdisciplinary approach and will develop rigorous methods at the intersection of mathematics, physics, statistics, and computer science, also making use of complex use cases from these domains.

In particular, the TP will address:

• Learning regimes with statistical-mechanics methods for the supervised and self-supervised case
• training dynamics of quantum neural networks with optimal mass transport theory
• development of foundational AI solutions through innovation in handling frontier’s scientific challenges
• design of explainability-oriented and trustworthy AI solutions in discriminative models
• generative models and reinforcement learning models
• rigorous trustworthy methods for automation of data handling in complex scientific/industrial environments, from real time to offline data analysis.