Project Details

Description

First-order optimization methods are methods of choice for large scale optimization problems. In general, such methods are poor at adapting to the geometry of the problem and can perform poorly on ill-conditioned problems. This projects investigates the class of so-called Bregman first-order optimization methods that are designed to better capture the problem geometry than traditional first-order methods do. We investigate theoretical properties as well as their application in applications domains such as machine learning training.
StatusActive
Effective start/end date2023/01/11 → …