Focus Period lund 2026
Thomas Pethick did his PhD with Prof. Volkan Cevher at École Polytechnique Fédérale de Lausanne (EPFL). His research is broadly concerned with optimization for machine learning with a focus on stabletraining of deep learning models.
Presenting: Training Deep Learning Models with Norm-Constrained LMOs
In this talk we will cover some non-Euclidean methods that have shown to be extremely efficient for training deep learning models. These developments are rooted in classical methods from optimization — in particular conditional gradient methods. We will see how to instantiate these methods for deep learning and discuss their favourable properties such as training stability, hyperparameter transfer, memory efficiency, scalability to large batches and their connection to weight decay. We close with some remarks regarding open problems.