Generative Priors for High-Dimensional BOED

Pretrained diffusion models as flexible priors for scalable BOED in non-Gaussian, high-dimensional inverse problems

Classical EIG estimation becomes computationally prohibitive when the unknown parameter is high-dimensional and the posterior is non-Gaussian. This ongoing work addresses this by using pretrained diffusion models as flexible priors for sequential BOED.

The key idea is to interpret posterior sampling as a guided diffusion trajectory, where the accumulated likelihood guidance along the latent path serves as a tractable surrogate for information gain. Building on mutual information estimation via score-based diffusion (Franzese et al., 2024), the MINDE-DPS estimator enables efficient EIG estimation alongside posterior sampling.

We validate the approach on source localization and Darcy flow parameter estimation, including high-dimensional PDE-constrained settings where classical methods struggle. For the function-space setting, we leverage FunDPS (Yao et al., 2025), enabling posterior sampling over PDE solution fields from extremely sparse observations. Ongoing work explores flow matching as a faster alternative for posterior sampling and EIG accumulation.

References

2025

  1. NeurIPS
    Guided Diffusion Sampling on Function Spaces with Applications to PDEs
    Jiachen Yao, Abbas Mammadov, Julius Berner, and 4 more authors
    arXiv preprint arXiv:2505.17004, 2025
    NeurIPS 2025

2024

  1. ICLR
    MINDE: Mutual Information Neural Diffusion Estimation
    Giulio Franzese, Mustapha Bounoua, and Pietro Michiardi
    In International Conference on Learning Representations, 2024