Joint Data and Computational Science seminar series
Title: Machine Learning and Uncertainty Quantification for Spatiotemporal Data Prediction
Speaker: Dr. Fei Miao and Dr. Huiqun Huang (Computer Science @UConn)
Accurate and efficient modeling of urban mobility—along with reliable trajectory prediction and object detection—is essential for the safety and resilience of intelligent transportation systems and smart cities. However, external dynamics such as weather, traffic interactions, and passenger behavior, as well as intrinsic data shifts, often cause distribution changes that degrade deep learning models and lead to overconfident predictions.
In this talk, I will introduce learning-based and statistical approaches to improve model robustness under such shifts. First, I present an attention-based framework for citywide anomaly prediction that captures spatio-temporal dependencies and quantifies the influence of urban mobility patterns on anomalies. Second, I discuss an extreme-aware model that decomposes mobility dynamics into regular and extreme components to enhance prediction under distribution shift. Third, I describe a conformal prediction and Gaussian process regression framework to quantify and reduce output uncertainty in trajectory forecasting models. Finally, I introduce an uncertainty-aware adversarial training method that improves the resilience of existing collaborative object detection models in autonomous driving by estimating and calibrating output uncertainty.
Textile 105A
: Zoom (available upon request)
Donghui Yan
dyan@umassd.edu
https://www.cscdr.umassd.edu/seminars