A0843
Title: Exploring Bayesian learning with heterogeneous image sources: From non-parametric models to deep learning architectures
Authors: Aaron Scheffler - University of California, San Francisco (United States)
Rajarshi Guhaniyogi - Texas A & M university (United States) [presenting]
Abstract: Various disciplines have recently encountered objects (such as tensors and networks) from multiple sources exhibiting diverse scales. These structured datasets offer a wealth of shared information among objects, which could potentially unveil crucial scientific insights through sophisticated analysis. Hierarchical Bayesian modeling approaches have the potential to effectively amalgamate information from heterogeneous objects using joint prior structures, thereby enabling comprehensive uncertainty estimation in inference. However, the adoption of this approach has been limited by the absence of robust Bayesian methods that yield precise inference, along with challenges in computation and theory. Innovative regression frameworks are introduced with an object outcome and heterogeneous object predictors encompassing both Bayesian non-parametric models and interpretable deep learning architecture to address these inferential challenges simultaneously. To validate the approach, empirical evidence is provided, particularly by analyzing high-impact multi-modal brain imaging datasets in collaboration with neuroscientists.
