Does Unsupervised Domain Adaptation Improve the Robustness of Amortized Bayesian Inference? A Systematic Evaluation

Neural networks are fragile when confronted with data that significantly deviates from their training distribution. This is true in particular for models trained on simulated data and deployed for inference on real-world observations. We formalize the targets of trustworthy inference and systematically test unsupervised domain adaptation (UDA) methods that align embeddings of simulated and real data. We find that UDA reduces noise-related errors, but can cause unexpected failures when the model's prior assumptions are wrong.