ECE Seminar Series - Burak Varici

Title: Causal Learning with Interventions: Estimation, Design, and Representation Learning

Abstract: Machine learning has seen great success in extracting useful patterns from data, but these patterns may not reflect the underlying mechanisms, limiting interpretability and reliability when conditions change. Causal models describe how a system generates data and responds to actions and environmental changes, providing a framework for connecting observations, predictions, and decisions. Interventions (changes to the underlying data-generating mechanisms) allow us to study these responses and distinguish between causal explanations that fit the same observational data.

In this talk, I will give a broad overview of my work organized around three questions: How can causal knowledge guide selecting effective interventions? How can interventions reveal causal structure when the underlying relationships and intervention targets are unknown? And how can we learn disentangled variables with causal semantics from high-dimensional observations, such as images? I will discuss these questions through work on causal bandits, intervention target estimation, structure learning from mixture data, and causal representation learning. Across these settings, I will highlight how modeling assumptions and the available interventions shape what can be learned and how efficiently, concluding with an illustration of causal representation learning in robotics.

Bio: Burak Varıcı is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Delaware. Before joining UD, he was a postdoctoral researcher at Carnegie Mellon University and received his Ph.D. in Electrical Engineering from Rensselaer Polytechnic Institute. His research lies at the intersection of causality and machine learning, focusing on the statistical and algorithmic foundations for learning reliable representations of complex data. His work spans causal representation learning, self-supervised learning, intervention design, and causal discovery. He is a recipient of the 2025 IEEE Signal Processing Society Best Ph.D. Dissertation Award.