Events

CDS Seminar - Beyond Invariance: Modeling Variation Across Environments

Speaker: Yuli Slavutsky (Columbia)

Location: 60 Fifth Avenue, Room 7th Floor Open Space

Date: Friday, October 30, 2026

Abstract: In multi-environment prediction, training data comes from several environments, and the goal is to predict in new ones. A prominent line of work seeks invariant representations by removing environment-specific variation. In this talk, we will examine a different strategy: treating environments as draws from a population, explicitly modeling their effects and marginalizing these effects when predicting in unseen environments.

We define predictive risk over a population of environments and show that its optimal predictor marginalizes over environment-specific conditional distributions. We then study the information retained by the representation in such models and examine the relationship to invariant representations. To implement this strategy, we develop neural generalized mixed-effects models that learn shared predictive structure and the distribution of environment-specific random effects. We show that these models extend to settings in which no invariant representation exists and outperform existing methods.

We then consider settings where environments affect the distribution of latent variables while the mechanisms generating observations and labels conditionally on those variables remain shared. We develop an empirical Bayes approach for marginalizing environment effects in these settings and show that the corresponding variational objective can be expressed as an environment-balancing objective. In scientific applications where such latent-variable structure is supported by domain knowledge, our approach outperforms existing environment-balancing methods, including invariant representation learning.

Based on joint work with Sebastian Salazar, Matthew Shen, Bohan Wu, and David M. Blei.

Bio: Yuli Slavutsky is a Founder’s Postdoctoral Research Scientist in the Department of Statistics at Columbia University. Her research develops Bayesian and variational methods for reliable prediction, with a focus on uncertainty quantification and robust representation learning. She received her PhD in Statistics and Data Science from the Hebrew University of Jerusalem, advised by Yuval Benjamini, and her master’s degree in Statistics also from the Hebrew University, advised by Or Zuk. Prior to her PhD, she worked in industry as a data scientist and research team leader.