Detecting Model Misspecification in Amortized Bayesian Inference with Neural Networks: An Extended Investigation

Generative AI enables efficient simulation-based inference (SBI) for complex models. But how faithful is such inference if the simulation represents reality somewhat inaccurately? We conceptualize the types of simulation gaps arising in SBI and investigate how neural network outputs gradually deteriorate as a consequence. To notify users about this problem, we propose a new measure that can be trained in an unsupervised way. We show that our measure warns users about suspicious outputs, raises an alarm when predictions are not trustworthy, and guides model designers in their search for better simulators.