Multidisciplinary Institute in Artificial intelligence

The chair develops novel optimization methods for machine learning, where one can exploit some structure in the learning pipeline. In particular, the omnipresence of discrete aspects –both of output objects, learned by models, and input objects, that we predict from– has arisen as a key property methods need to capture. Incorporating discrete structures when building and training models opens a broad set of challenges due to harder optimization problems. Among our end-goals, we want to build robustness against various forms of perturbations into such learning models by leveraging the additional structure.