Pierfrancesco Urbani
CNRS researcher at the Institute of Theoretical Physics, CEA Saclay, France
IAS SEMINAR #08/2026
Turin / OGR / Sala Duomo / June 18, 2026 / 04:00 – 06:00 pm CEST
Separation of timescales controls feature learning and overfitting in large neural networks
EVENT&WEBINAR
Understanding how large overparameterized networks can extract relevant features from meaningful data while simultaneously be so expressive to perfectly interpolate pure noise is a conceptual problem central in machine learning. To address this, we analyze the out-of-equilibrium, non-linear, high-dimensional training dynamics of overparametrized two-layer neural networks using dynamical mean field theory.
Our findings reveal a separation of timescales in training, leading to several key observations:
- A slow timescale linked to the growth in model complexity;
- An inductive bias towards low complexity when the initial model complexity is sufficiently small;
- A dynamical decoupling between feature learning and overfitting phases;
- A non-monotonic trend in test error, characterized by a “feature unlearning” regime at later stages of training. Joint work with Andrea Montanari.
PIERFRANCESCO URBANI
Pierfrancesco Urbani is a CNRS Researcher working at the Institute of Theoretical Physics (CEA Saclay), France, and a Part-time Professor at École Polytechnique. His research focuses on statistical physics applied to complex systems, particularly the physics of amorphous solids. He has developed a theory of the glass phase in infinite spatial dimensions, identified new types of phase transitions, exactly characterized the jamming transition of granular materials, and formulated a theory of amorphous solids under load. With Giorgio Parisi and Francesco Zamponi, he co-authored the monograph Theory of Simple Glasses (Cambridge University Press, 2020).
In recent years, his work has expanded into high-dimensional inference and optimization. He developed a theory of continuous constraint satisfaction problems and analyzed gradient-based optimization algorithms using dynamical mean field theory. Key contributions include the study of Canyon Landscapes as models of overparameterized loss landscapes in machine learning and solving the training dynamics of overparameterized neural networks. Currently, he focuses on the learning dynamics of recurrent neural networks and nonlinear dynamical systems, exploring problems at the intersection of statistical physics, computational neuroscience, and machine learning.
In collaboration with:
Register for In-Person Attendance
Joining us in person offers the opportunity for direct networking and interaction with both speakers and participants.
Please complete the form below to reserve your seat.
By clicking, you will be automatically registered for the in-person event.
Note: If you see a message saying you’re already subscribed, please ignore it — your in-person registration will still be confirmed.
Please note that, by attending, you implicitly authorize AI4I to film and photograph the audience during the seminar. By submitting your registration, you acknowledge and accept the AI4I Data Privacy Policy. Seats are limited: if you are unable to attend, we kindly ask you to inform us as soon as possible at ias@ai4i.it
LINK ZOOM
