A1175
Title: Unraveling complex relationships between misaligned images with additive neural network Gaussian processes
Authors: Rene Gutierrez - University of Texas at El Paso (United States) [presenting]
Rajarshi Guhaniyogi - Duke University (United States)
Aaron Scheffler - UCLA (United States)
Abstract: The presentation explores relationships among images of varying scales, resolutions, and shapes, addressing a key issue in image-based data. The focus is on misalignment from data captured at different scales: upper-level images segmented into regions and lower-level images segmented into sub-regions. Within a regression framework, the response region and a subset of predictor images are defined at the lower scale, with an additional predictor image derived from a network where nodes represent upper-level regions. A regression framework that captures non-linear effects of network and lower-level predictors on responses is introduced, using a Gaussian process (GP) prior with a neural network-based covariance function (NN-GP prior). This framework merges Bayesian GP regression's uncertainty quantification with neural networks' predictive power, offering a flexible non-linear regression approach. The method enables scalable computation, captures intricate characteristics often missed by local spatial smoothing methods, and improves robustness by treating lower-scale image sub-regions as effective samples. Simulation studies show that the approach outperforms existing methods in predictive inference for response images.
