Amortized Bayesian inference is a simulation-based technique for estimating posteriors when likelihoods are costly or impossible to evaluate directly. The approach has a known weakness: in high-dimensional models, learning the full joint posterior becomes very difficult.
The new arXiv paper, "Semi-Amortized Bayesian Inference with Reusable Local Posteriors," takes a semi-amortized route. Rather than treating the whole posterior as a single learning problem, the title and abstract indicate that the method breaks the problem into local posteriors that can be reused. The abstract highlights the challenge of full-joint-posterior learning as the central motivation.
Because the available source is only the abstract, the article can describe the proposal at a high level. No experimental results or comparisons are included in the provided text.