Analyzing Inverse Problems in Natural Science using Invertible Neural Networks

Ullrich Koethe, HCI

November 20, 2019


Uncertainty quantification is a hot topic in neural network research. This talk will focus on inverse problems, where high uncertainty arises from the inherent ambiguities of ill-posed inverse processes. This type of problem is ubiquitous in natural sciences, and existing approaches are either very expensive or suffer from drastic approximations. The talk presents a new class of invertible neural networks that generalize established Bayesian approaches from the linear to the non-linear setting. These networks work equally well in the forward as well as the inverse direction and thus enable new training and approximation methods, which become asymptotically exact in the perfect convergence limit. A variety of promising results from medical imaging, computer vision, and environmental physics demonstrate the practical utility of the new method.


Prof. Ullrich Koethe is heading the newly founded group on “Explainable Machine Learning” at the Visual Learning Lab Heidelberg. He is interested in the connection between machine learning and its applications in image analysis, medicine, and natural sciences. Explainable learning shall open-up the blackbox of successful learning algorithms, in particular neural networks, to provide insight rather than mere numbers. In addition, he is interested in generic software bringing state-of-the-art algorithms to the end user and maintain the VIGRA image analysis library.

Event Info

The event will take place on Wednesday, 20 November, 2019 at 7:00pm at the DKFZ Communication Center (K1+K2), Im Neuenheimer Feld 280. Drinks and snacks will be provided, courtesy of the Division of Medical Image Computing at DKFZ. Kindly help us plan ahead by registering for the event on our meetup page.