Date: Tue. July 1, 2025 (12:30H)
Event: DeustoCCM Seminar

Title: Universal approximation and convexified training in neural networks
Speaker: Prof. Dr. Kang Liu
Affiliation: Université Bourgogne Europe

Abstract. The universal approximation property (UAP) of neural networks (NNs) provides a theoretical foundation for the modern era of artificial intelligence. In this presentation, we revisit several key UAP results for NNs and extend them to dynamical systems through the framework of semi-autonomous neural ordinary differential equations (SA-NODEs). Although the approximation error can be made arbitrarily small, training NNs involves solving high-dimensional, non-convex optimization problems. By leveraging mean-field techniques, we convexify the training process, offering valuable insights into practical aspects such as generalization bounds, hyperparameter selection, and scaling laws.

WHERE
University of Deusto, DeustoTech.
Room: Logistar
Unibertsitate Etorb., 24, 48007 Bilbao, Bizkaia