Ullrich Köthe (Université de Heidelberg) is leading the DATAIA seminar of September on « Solving inverse problems with invertible neural networks ».

Interpretable models are a hot topic in neural network research. This talk will focus on inverse problems, where one wants to infer backwards from observations to the hidden characteristics of a system. I will focus on three aspects of interpretability: reliable uncertainty quantification, outlier detection, and disentanglement into meaningful features. It turns out that invertible neural networks -- networks that work equally well in the forward and inverse direction -- are great tools for that kind of analysis: They act as non-linear generalizations of classical methods like PCA and ICA. Examples from physics, medicine, and computer vision demonstrate the practical utility of the new method.