The Gaussian Latent Machine: Efficient Prior and Posterior Sampling for Inverse Problems

A team of researchers from institutions including the University of Vienna and EPFL has published a new paper on arXiv titled "The Gaussian Latent Machine: Efficient Prior and Posterior Sampling for Inverse Problems." The work introduces a novel latent variable model designed to streamline a core challenge in Bayesian imaging: sampling from complex probability distributions. These distributions, often formulated as a "product of experts," are common when reconstructing images from noisy or incomplete data, such as in MRI scans or astronomical imaging. The Gaussian Latent Machine elegantly lifts these difficult sampling problems into a latent space where operations become more tractable.

This framework unifies and generalizes many existing sampling algorithms under one roof. Its most significant contribution is a general, highly efficient two-block Gibbs sampling approach. In specific cases, the model even allows for direct sampling, bypassing iterative methods entirely. The authors back their theoretical work with detailed numerical experiments, demonstrating the method's effectiveness across a wide range of standard Bayesian imaging problems. The result is a more robust and computationally efficient toolkit for researchers and engineers working on inverse problems where accurate uncertainty quantification is critical.