Designing Priors for Bayesian Neural Nets to Enhance Probabilistic Predictive Models in Engineering

This seminar presents Dr. Audrey Olivier's research on designing priors for Bayesian neural networks to enhance probabilistic predictive models in engineering applications. Learn how the integration of data mining and physics-based modeling can improve the design, monitoring, and optimization of engineering systems. Discover enhanced algorithms based on ensembling with anchoring for approximate Bayesian learning of neural networks, with special focus on prior design that incorporates knowledge from low-fidelity models and accounts for low-rank correlations between neural network weights. The presentation showcases applications across civil engineering domains, including surrogate training for materials and structural modeling, contingency analysis in power grid systems, and ambulance travel time prediction in urban networks. Dr. Olivier, an Assistant Professor at USC's Sonny Astani Department of Civil and Environmental Engineering, addresses the unique challenges of applying machine learning to engineering contexts where datasets are often noisy, sparse, and imbalanced, emphasizing the importance of robust uncertainty quantification for high-consequence decision-making.