Reproduce “A Black Box Variational Inference Scheme for Inverse Problems with Demanding Physics-Based Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
G. Robalo Rei, C.P. Schmidt, J. Nitzler, M. Dinkel, W. A. Wall
Researchers at the Technical University of Munich developed Adaptive Batch-sized Regulated Importance Sampling (ABRIS) Black Box Variational Inference (BBVI), a method that significantly reduces the number of expensive forward model evaluations required for Bayesian inverse problems. This approach, which reuses past computations and adaptively triggers new samples, enabled robust uncertainty quantification for complex black-box physics models, often with orders of magnitude fewer model calls than traditional methods.
Key insights
- Importance sampling allows for the efficient reuse of likelihood evaluations from past variational distributions, drastically reducing the need for new, costly forward model runs.
- Adaptive criteria can intelligently identify when the existing sample pool becomes insufficient for accurate gradient estimation, enabling numerous 'sampling-free' iterations where only cheap re-weighting occurs.
- The batch-sequential nature of new model evaluations inherently facilitates parallel processing, making the method efficient on High-Performance Computing (HPC) infrastructures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.