Events Calendar

AI Seminar- Designing Priors for Bayesian Neural Networks to Enhance Probabilistic Predictive Modeling in Engineering Applications

Speaker: Audrey Olivier, USC

Talk Title: Designing Priors for Bayesian Neural Networks to Enhance Probabilistic Predictive Modeling in Engineering Applications

Series: AI Seminar

Abstract: The conjuncton of data mining and physics-based modeling holds great potential to help design, monitor and optimize engineering systems. Efficient ML algorithms can uncover patterns from data to learn missing physics, detect abnormal behaviors and identify damaged systems, or serve as surrogates of complex mechanistic models, enabling real-time analysis or integration within optimization frameworks. However, the use of ML for engineering applications and high-consequence decision-making presents unique challenges. Engineering datasets are often noisy, sparse and imbalanced, due to the inherent randomness of the underlying physical processes and constraints on data collection. Whenever possible, ML predictors must assimilate physics-based knowledge and intuitions to improve accuracy and generalization away from training data. Most importantly, ML models must embed robust and reliable prediction of uncertainties to improve trustworthiness for high-consequence decision-making. Framing ML training within a Bayesian inference framework allows for a robust quantification of both aleatory and epistemic uncertainties that arise from data inadequacies, integration of physics intuitions through prior design, and assessment of the model’s confidence in its predictions. However, due to the high-dimensionality and non-physicality of parameters that characterize typical ML models such as neural networks, application of Bayesian methods in this context raises several technical challenges, from prior and likelihood design to posterior inference. This talk will introduce enhanced algorithms based on ensembling with anchoring for approximate Bayesian learning of neural networks. We will demonstrate the importance of carefully designing the prior, integrating knowledge from low-fidelity models via ensemble pre-training and designing parameter-space prior densities that account for low-rank correlations between neural network weights. The talk will illustrate the benefits of these methods through a variety of example applications in civil engineering, from surrogate training to accelerate materials and structural modeling, contingency analysis in power grid systems, or ambulance travel time prediction in a dense urban network to help optimize emergency medical services.

Biography: Dr. Olivier holds a Diplôme d’Ingénieur from École Centrale de Nantes, France, and a Ph.D. in Civil Engineering and Engineering Mechanics from Columbia University, USA. She held a postdoctoral appointment at Johns Hopkins University before joining the Sonny Astani Department of Civil and Environmental Engineering at the University of Southern California as an Assistant Professor in Fall 2021. Dr. Olivier’s research aims to predict and monitor civil infrastructure systems behavior under uncertainty, by combining innovations in probabilistic data analytics and mechanistic modeling. Applications span various scales, from systems to structures to materials, and include development of adaptive Bayesian filters for identification of dynamical structural systems, probabilistic surrogate models to accelerate multi-scale materials simulations or Bayesian graph neural networks for contingency analysis of power grids.

Host: Eric Boxer and Justina Gilleland

More Info: https://www.isi.edu/events/5453/designing-priors-for-bayesian-neural-networks-to-enhance-probabilistic-predictive-modeling-in-engineering-applications/

Webcast: https://usc.zoom.us/j/94409584905?pwd=Sm5LVkd0bndUdEluM3piK0NWTUQrUT09

Location: Information Science Institute (ISI) - Conf Rm#1135 and Virtual

WebCast Link: https://usc.zoom.us/j/94409584905?pwd=Sm5LVkd0bndUdEluM3piK0NWTUQrUT09

Audiences: Everyone Is Invited

Contact: Pete Zamar

Event Link: https://www.isi.edu/events/5453/designing-priors-for-bayesian-neural-networks-to-enhance-probabilistic-predictive-modeling-in-engineering-applications/

This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.