Amortized Inference
Overview
Amortized inference learns a reusable mapping from observations and model information to posterior estimates, predictions, or samples, shifting computation from each new case to training. Common approaches include simulation-based neural posterior estimation, variational encoders and normalizing flows, structured message passing, and transformer-based prior-fitted or in-context models. These methods are used in settings such as inverse problems, state-space models, survival analysis, and hierarchical statistics, often with MCMC or benchmark posteriors used to assess accuracy and calibration. The central tension is that fast single-pass inference can lose accuracy or uncertainty calibration on rare cases, shifted data, or complex posteriors, motivating test-time refinement and stronger evaluations of generalization.
Latest papers
87 papers17 papers in the last four weeks, up 113% on the four weeks before. 0.1% of all new papers.