Building trustworthy Al systems requires models that know how much to trust their own predictions. This project develops methods for credibility estimation, scoring each prediction by how likely it is to be correct, as a foundation for more reliable machine learning.
The Bayesian framework offers a principled way to model this kind of uncertainty, but its computational cost has made it impractical for modern deep learning, where models now routinely reach billions of parameters. As a result, Bayesian methods have been sidelined even as the need for calibrated uncertainty has grown more urgent.
This research bridges that gap by developing tractable approximations to Bayesian inference, grounded in geometric principles drawn from information geometry, randomized linear algebra, and numerical optimization. The aim is to scale these techniques to neural networks in the hundreds-of-millions-of-parameters range, covering a wide range of modern deep learning architectures.