Biometrics, 80(2)

We build a Bayesian model that combines CNNs with GLMs to make predictions from images and other types of data simultaneously. By using Monte Carlo (MC) dropout approximation, we capture uncertainty in the image features and reflect it in the final prediction. It’s simple, fast, and gives you reliable results with uncertainty built in!

Journal of Geophysical Research - Atmospherics  

Wildfire smoke makes it hard to measure AOD in California, leaving lots of missing data. We use a Bayesian CNN with spatiotemporal basis functions to better fill in the gaps and capture space-time patterns. This helps improve predictions while accounting for uncertainty in a realistic way. This work is done with researchers from the Center for Spatial Information Science and Systems at George Mason University.

Now available on arXiv

We propose a scalable generative model for high-dimensional neuroimaging data that integrates spatially-varying cortical features and a network-valued coactivation structure via deep neural networks. The model captures complex non-linear and multi-scale spatial associations, enforces spatial smoothness, and accounts for subject-level heterogeneity. By incorporating Monte Carlo dropout, it provides uncertainty-aware prediction and interpretable effect maps, forming one of the first Explainable AI (XAI) frameworks for spatial-network-valued imaging data. This model achieves state-of-the-art performance in predictive accuracy, uncertainty quantification, and computational efficiency without requiring extensive pre-processing or dimension reduction.

Now available on arXiv

We introduce DeepSurrogate, an AI-powered surrogate model for functional outputs with vector-valued inputs, designed to emulate high-fidelity simulators efficiently. The model represents input–output relationships as spatially-indexed functions, decomposed into input-driven nonlinear effects and spatial dependencies, both learned via deep neural networks. This architecture captures complex spatial correlation while maintaining scalability. Predictive uncertainty is quantified using Monte Carlo dropout, enhancing interpretability. DeepSurrogate is highly efficient, capable of training on tens of thousands of spatial locations within minutes, and demonstrates strong performance across synthetic and real-world scientific simulations such as the SLOSH storm surge model.

Now available on arXiv

Understanding relationships across multiple imaging modalities is central to neuroimaging research. We introduce the Integrative Variational Autoencoder (InVA), the first hierarchical VAE framework for image-on-image regression in multimodal neuroimaging. Unlike standard VAEs, which are not designed for predictive integration across modalities, InVA models outcome images as functions of both shared and modality-specific features. This flexible, data-driven approach avoids rigid assumptions of classical tensor regression and outperforms conventional VAEs and nonlinear models such as BART. As a key application, InVA accurately predicts costly PET scans from structural MRI, offering an efficient and powerful tool for multimodal neuroimaging.