Learning and Leveraging Rich Priors for Factorization Problems

Research Project, 2020 –

In this project we are interested in developing methods that combine traditional (parametric) mathematicalformulations induced by domain expertise with (non-parametric) models learned from examples. Parametricmodels inject domain knowledge into learning-based approaches and have therefore the potential to massivelyreduce the necessary amount of training data. Additionally, the output can be constrained e.g. to be physicallyplausible, which is difficult to guarantee with pure learning-based architectures. At the same time, being able toincorporate e.g. learned priors has the potential to regularize problems where a physical model is not sufficientto guarantee a well posed formulation. From a theoretical point of view we are interested in results thatcharacterize formulations in terms of their expressiveness and generalization as well as developing efficientinference approaches.