Conditional Invertible Generative Models for Supervised Problems

Invertible Neural Networks and normalizing flows have various attractive properties compared to types of generative models in theory. However, they are rarely useful for real applications due to their unguided outputs. In this work, we present three new methods that extend this standard setting, which we generally term "generative invertible models". These new methods allow leveraging the theoretical and practical benefits of invertible neural networks to solve supervised problems in new ways, demonstrated with multiple real-world applications from different branches of science. The key finding is that our generative invertible approaches enhance aspects of trustworthiness in comparison to conventional approaches, such as uncertainty quantification, explainability, and handling of outlier data.