S'abonner à l'agenda

S'abonner au séminaire

Ajouter cette URL aux abonnements de votre calendrier électronique :
https://irma.math.unistra.fr/cal/seminaire-equations-aux-derivees-partielles.ics
Fermer

Séminaire Equations aux dérivées partielles

organisé par l'équipe Modélisation et contrôle

  • Ivan Dokmanić

    A spring-block theory of feature learning in deep neural networks

    3 décembre 2024 - 14:00Salle de conférences IRMA

    A central question in deep learning is how deep neural networks (DNNs) learn features. DNN layers progressively collapse data into a regular low-dimensional geometry. This collective effect of nonlinearity, noise, learning rate, width, depth, and numerous other parameters, has eluded first-principles theories which are built from microscopic neuronal dynamics. We discovered a noise–nonlinearity phase diagram that highlights where shallow or deep layers learn features more effectively. I will describe a macroscopic mechanical theory of feature learning that accurately reproduces this phase diagram, offering a clear intuition for why and how some DNNs are ``lazy'' and some are ``active'', and relating the distribution of feature learning over layers with test accuracy. Joint work with Cheng Shi and Liming Pan.

  • Yvonne Alama Bronsard

    TBA

    16 janvier 2025 - 14:00Salle de conférences IRMA

    TBA

  • Elise Grosjean

    TBA

    4 février 2025 - 14:00Salle de conférences IRMA

  • Simon Schneider

    TBA

    25 février 2025 - 14:00A confirmer