Probabilistic Bayesian optimal experimental design using conditional normalizing flows
This work addresses efficiency and scalability issues in Bayesian OED for practical applications like medical imaging, though it appears incremental as it builds on existing methods with specific improvements.
The paper tackles the computationally challenging Bayesian optimal experimental design (OED) problem by proposing a novel joint optimization approach using conditional normalizing flows and a probabilistic formulation for binary designs, demonstrating its performance on a high-dimensional MRI data acquisition problem with parameters and observations in the hundreds of thousands.
Bayesian optimal experimental design (OED) seeks to conduct the most informative experiment under budget constraints to update the prior knowledge of a system to its posterior from the experimental data in a Bayesian framework. Such problems are computationally challenging because of (1) expensive and repeated evaluation of some optimality criterion that typically involves a double integration with respect to both the system parameters and the experimental data, (2) suffering from the curse-of-dimensionality when the system parameters and design variables are high-dimensional, (3) the optimization is combinatorial and highly non-convex if the design variables are binary, often leading to non-robust designs. To make the solution of the Bayesian OED problem efficient, scalable, and robust for practical applications, we propose a novel joint optimization approach. This approach performs simultaneous (1) training of a scalable conditional normalizing flow (CNF) to efficiently maximize the expected information gain (EIG) of a jointly learned experimental design (2) optimization of a probabilistic formulation of the binary experimental design with a Bernoulli distribution. We demonstrate the performance of our proposed method for a practical MRI data acquisition problem, one of the most challenging Bayesian OED problems that has high-dimensional (320 $\times$ 320) parameters at high image resolution, high-dimensional (640 $\times$ 386) observations, and binary mask designs to select the most informative observations.