Interdisciplinary Institute in Artificial Intelligence

The chair develops novel optimization methods for machine learning, where one can exploit certain structures in the learning pipeline. In particular, the ubiquity of discrete aspects—both in the output objects learned by models and in the input objects from which we make predictions—has emerged as a key property that methods must capture. Incorporating discrete structures when building and training models presents a wide range of challenges due to more complex optimization problems. Among our ultimate goals is to build robustness against various forms of perturbations into such learning models by leveraging this additional structure.