[Dottorcomp] Fwd: Annuncio di Seminario Matematico e Fisico

[Stefano Lisini]

Cari colleghi,

inoltro il seguente annuncio del Seminario Matematico e Fisico, Milano.

Cordiali saluti,
Stefano Lisini.

_______________________________

Charalambos Makridakis (IACM e Università del Sussex)

Giovedì 14/05/2026, aula Mn, via Mangiagalli 32

Titolo
Understanding Machine Learning through Mathematics: A Variational Approach

Abstract
In recent years, substantial progress has been made in understanding the
approximation properties of neural networks, with foundational results
showing that a range of architectures can efficiently approximate both
smooth and singular functions. However, these advances do not by themselves
explain the behaviour of machine learning algorithms, especially in
applications such as PDE solvers, uncertainty quantification, and
generative models.

This talk presents a mathematical framework for analysing ML algorithms
through their underlying variational principles and loss functionals. By
studying function learning, PDE solvers, and probability-measure
approximation from this perspective, we relate classical notions of
numerical stability to the properties of the associated variational
problems. This approach yields new insights into stability and convergence
and motivates the development of new algorithms.

________________________________