Aller au contenu. | Aller à la navigation

Outils personnels

Navigation

UMR 5672

logo de l'ENS de Lyon
logo du CNRS
Vous êtes ici : Accueil / Séminaires / Machine Learning and Signal Processing / A physicist's approach to AI systems

A physicist's approach to AI systems

Marco Biroli (postdoctoral researcher at the University of Chicago)
Quand ? Le 17/11/2026,
de 13:00 à 14:00
Où ? M7 101
Ajouter un événement au calendrier vCal
iCal
Abstract: 
Modern AI systems have reached remarkable performance, yet we still struggle to explain how they behave. Most attempts to understand them are either empirical (run models, fit curves) or axiomatic (prove theorems about idealized scenarios). In this talk I will show how a physicist’s approach, postulating effective theories distilled from experimental phenomenology, can lead to predictive models that give insight into the inner workings of these networks.
 
I will showcase two examples. First, I will show how concepts from machine learning (variational autoencoders), mathematics (conditional independence) and physics (mean-field approximations) are intertwined. Unwinding these relationships gives a sharp criterion for when such models can succeed or must fail, and lets us read the physics a model has learned directly off its trained weights. Second, I will show how the same approach captures the failures of modern large language models (LLMs) with the statistical physicist’s usual toolkit, extreme-value statistics and large-deviation theory: a five-parameter model that predicts how an adversarial attack scales far beyond the regime it was fitted in.
 
I will end on a more philosophical outlook on the place of AI in science, and on the path ahead.