JANA: Jointly Amortized Neural Approximation of Complex Bayesian Models - 2023

Title : JANA: Jointly Amortized Neural Approximation of Complex Bayesian Models Author(s): Radev, Stefan T. and Schmitt, Marvin and Pratz, Valentin and Picchini, Umberto and Köthe, Ullrich and Bürkner, Paul-Christian Link(s) : http://arxiv.org/abs/2302.09125

Proposed a jointly amortized neural approximation (JANA) of intractable likelihood functions and posterior densities in Bayesian surrogate modelling and simulation-based inference. They train:

1. A summary network to compress each data point/time series into embedding vectors.
2. A posterior network to learn an amortized approximate posterior.
3. A likelihood network to learn an amortized approximate likelihood.

When optimizing the training objective (Eq 9), the networks can then be used for tasks such as posterior predictive estimation and marginal likelihood estimation.