Please join us online on Wednesday, October 14, at 1:30 pm Central Time, for a talk by Dr. Yuqi Gu. She will present “Statistical Foundations of Identifiable Latent Representations: From Deep Generative Models to Causal Structure." For access to this seminar, contact series administrator Cierra Streeter.
is Assistant Professor of Statistics and a member of the Data Science Institute at Columbia University. She develops statistical theory and methodology for learning interpretable latent structure from complex, high-dimensional data, with a particular emphasis on identifiability, high-dimensional inference, and representation learning. Her recent work studies identifiable deep generative models, causal representation learning, and statistical measurement and evaluation of modern AI systems. She received her PhD in Statistics from the University of Michigan in 2020 and was a postdoctoral researcher at Duke University before joining Columbia in 2021.Ìý
Abstract
Identifiability is a fundamental requirement for interpretable representation learning: if very different latent explanations can generate the same observed data, the meaning of a learned representation, and any scientific conclusions built on it, can be unstable. Modern representation learning therefore raises a basic statistical question: when is hidden structure uniquely determined by complex, high-dimensional observations? In this talk, Dr. Gu will develop a statistical approach to this question using discrete latent representations. She will first introduce , a class of deep generative models whose layered latent structure can be identified under interpretable conditions, while retaining flexible modeling and scalable estimation. Applications to images, text data, and multimodal educational assessment data illustrate how the learned representations can capture meaningful hierarchical structure. She will then ask a harder question: once latent factors are identifiable, can we also learn how they causally influence one another? Her team’s framework shows that, under suitable structure, noisy and heterogeneous observations can reveal both how latent factors generate the data and the causal structure among those factors that can be recovered from observational data alone. Together, these results provide a path from flexible representation learning to statistically grounded interpretation, causal discovery, and ultimately intervention.