Uncertainty-Aware Distribution-to-Distribution Flow Matching for Scientific Imaging
Researchers have developed a new framework called Bayesian Stochastic Flow Matching (BSFM) to improve the reliability and accountability of generative models used in scientific imaging. This approach builds upon Stochastic Flow Matching (SFM) by incorporating uncertainty quantification, allowing for better generalization across different experimental conditions and the detection of unreliable predictions. Experiments on cellular imaging and fMRI data demonstrate that BSFM effectively provides anomaly scores for detecting out-of-distribution cases within practical sampling budgets. AI
IMPACT Enhances trustworthiness of AI models in scientific applications by quantifying uncertainty and detecting unreliable outputs.