Please note this event occurred in the past.
October 23, 2026 2:00 pm - 3:00 pm ET
Statistics and Data Science Seminar Series
LGRT 1685

Title: Representation Learning and Statistical Generative Modeling for Distributions
 

Abstract: Modern instruments such as wearables, imaging devices, and geospatial sensors generate rich data streams of thousands to millions of measurements. Collapsing these streams into simple summaries like means can obscure meaningful structure, whereas retaining the full distribution of measurements preserves much of the information these technologies provide.

 
In this talk, I will present a strategy for modeling distributions of wearable recordings, using physical activity data from the National Health and Nutrition Examination Survey (NHANES) as a motivating case study. This strategy is one instance of a more general approach to analyzing complex objects: statistical generative modeling (SGM). The core idea is to engineer a multi-layer encoder that maps each object to a latent embedding whose statistically convenient properties, such as compactness, near-losslessness, and approximate Gaussianity, make it amenable to principled inference with familiar statistical models. Each encoding layer is a mathematical, statistical, or AI/ML transform designed to impart one or more of these properties. A paired decoder maps embeddings back to object space, so latent modeling and inference can be interpreted in terms of the original objects. Because the encode-model-decode pipeline is generative, it also enables a powerful model diagnostic: simulate synthetic objects from the fitted model and check whether they resemble the real data.
 
For NHANES activity data, we represent each participant's week of minute-level recordings by a quantile function and build a three-layer encoder that achieves near-lossless reconstruction in just five dimensions. We demonstrate the full SGM workflow through two analyses, group comparison and semiparametric regression, illustrating how familiar statistical tools become available for distributional outcomes. 

 


Short bio: Dr. Kyle Stanley is a postdoc in the Department of Biostatistics, Epidemiology, and Informatics at the University of Pennsylvania, working with Jeffrey Morris. His research focuses on functional and distributional data analysis, representation learning, and generative modeling, with applications to wearable devices, imaging, and other complex object data. He received his PhD in Statistics from Penn State.