Dear all / Bonjour à tous,

We are happy to invite you to the PhD Defense of Charles Guille-Escuret on July 28th at 10 am (hybrid mode).

Vous êtes cordialement invité.e.s à la défense de doctorat de Charles Guille-Escuret, le 28 juillet à 10h00 (mode hybride).

Jury

Résumé:

This work argues that the significant gap between optimization theory and practice in AI stems from a poor understanding of the objective functions being optimized. To build a more relevant theory, we must first better characterize the class of functions encountered in modern machine learning.

The research critiques standard theoretical assumptions, showing they can lead to misleading conclusions about algorithm performance. By exploring alternative, empirically-grounded properties, it demonstrates that theoretical outcomes—such as the benefits of acceleration—are highly sensitive to these foundational choices.

An experimental study reveals that practical deep learning optimization exhibits consistent geometric properties not captured by standard models. For instance, the success of the popular Adam optimizer is shown to be dependent on the coordinate system, a crucial detail that most theoretical frameworks cannot capture.

Overall, this research highlights the limitations of prevailing theory and makes the case that future progress requires a deeper, more empirically-informed characterization of objective functions in deep learning.