Reproduce “Conformalized Generative Bayesian Imaging: An Uncertainty Quantification Framework for Computational Imaging” on OpenResearch

About this paper

Canberk Ekmekci, Mujdat Cetin

This framework combines generative models with Bayesian deep learning and conformal prediction to provide comprehensive uncertainty quantification for computational imaging inverse problems. It jointly quantifies both aleatoric and epistemic uncertainties, offering statistically guaranteed prediction sets and demonstrating improved reconstruction quality across various imaging tasks.

Key insights

• Deep ensembling effectively approximates the posterior over generative model parameters, enabling the decomposition and quantification of both aleatoric and epistemic uncertainties.
• Epistemic uncertainty naturally decreases with more training data and increases for out-of-distribution inputs or anomalies, making it a valuable indicator for model trustworthiness.
• Conformal prediction can be seamlessly integrated with complex deep learning UQ methods to provide rigorous statistical coverage guarantees, transforming heuristic uncertainty estimates into verifiable ones.

Summary by alphaXiv.