A1119
Title: Latent representation and modeling of structured functional data
Authors: Edward Gunning - University of Pennsylvania (United States) [presenting]
Giles Hooker - University of Pennsylvania (United States)
Jeffrey Morris - University of Pennsylvania (United States)
Ye Emma Zohner - Rice University (United States)
Abstract: Functional data, in particular, those observed on two- or higher-dimensional domains (e.g., images, surfaces), typically present as structured, ultra-high-dimensional vectors in practice. Dimension reduction of these data is typically used to transform them to a lower-dimensional space of features that facilitates conventional statistical modeling. Traditionally, FDA methods have relied on linear methods for dimension reduction (e.g., basis function representations, PCA), but more modern machine learning approaches (e.g., autoencoders) offer viable alternatives. The contributions are twofold: first, a flexible evaluation framework is developed to select an optimal (linear or nonlinear) latent representation of a functional dataset. Subsequently, it is demonstrated how, when a nonlinear representation is selected, it can be embedded within the conventional Bayesian functional mixed model framework of a past study.
