ORNL researchers developed a stochastic approximate gradient ascent method to reduce posterior uncertainty in Bayesian experimental design involving implicit models.
- introduces a neural-network-based mutual information estimator into the objective function to replace the computationally intractable likelihood function, and
- can simultaneously train the mutual information estimator and search for an optimal design.
- achieves designs with significantly improved confidence, i.e., small uncertainty of the posterior distribution when compared to the state-of-the-art Bayesian optimization (BO) method, and
- avoids intractable overhead cost of BO in solving high-dimensional problems.
PI: Guannan Zhang
Publication: J. Zhang, S. Bi, and G. Zhang, A hybrid gradient method to designing Bayesian experiments for implicit models, NeurIPS Workshop Proceeding on Machine Learning and the Physical Sciences, Dec. 2020. (PDF file).
ASCR Program/Facility: DOE ASCR Applied Math
Funding: DOE ASCR Applied Math (UQ4ML), and the ORNL AI Initiative