Optimal Experimental Design for Bayesian Inverse Problems Using Energy-Based Couplings

Abstract

Bayesian Experimental Design (BED) is a robust model-based framework for optimising experiments but faces significant computational barriers, especially in the setting of inverse problems for partial differential equations (PDEs). In this paper, we propose a novel approach, modelling the joint posterior distribution with an energy-based model, trained on simulation data. Unlike existing simulation-based inference approaches, we leverage implicit neural representations to learn a functional representation of parameters and data. This is used as a resolution-independent plug-and-play surrogate for the posterior, which can be conditioned over any set of design-points, permitting an efficient approach to BED.

Cite

Text

Encinar et al. "Optimal Experimental Design for Bayesian Inverse Problems Using Energy-Based Couplings." ICLR 2024 Workshops: AI4DiffEqtnsInSci, 2024.

Markdown

[Encinar et al. "Optimal Experimental Design for Bayesian Inverse Problems Using Energy-Based Couplings." ICLR 2024 Workshops: AI4DiffEqtnsInSci, 2024.](https://mlanthology.org/iclrw/2024/encinar2024iclrw-optimal/)

BibTeX

@inproceedings{encinar2024iclrw-optimal,
  title     = {{Optimal Experimental Design for Bayesian Inverse Problems Using Energy-Based Couplings}},
  author    = {Encinar, Paula Cordero and Schröder, Tobias and Duncan, Andrew B.},
  booktitle = {ICLR 2024 Workshops: AI4DiffEqtnsInSci},
  year      = {2024},
  url       = {https://mlanthology.org/iclrw/2024/encinar2024iclrw-optimal/}
}