Interpretability and causality are key desiderata in modern machine learning systems. Graphical models, and more specifically directed acyclic graphs (DAGs, a.k.a. Bayesian networks), serve as a well-established tool for expressing interpretable causal relationships. However, the task of estimating DAG structures from data poses a significant challenge, given its inherently complex combinatorial nature, and traditional approaches rely on various local heuristics. In this presentation, I will explore an innovative approach that fundamentally redefines the causal discovery problem as a smooth but nonconvex optimization problem that avoids combinatorial constraints entirely. Following an overview of this framework and recent advancements in understanding its properties, I will delve into recent progress in directly learning causal shifts and invariances from multiple datasets, bypassing the estimation of causal structures. Finally, I will provide an overview of open problems and my future research plans for causal machine learning.

Dr. Kevin Bello is an NSF Computing Innovation Fellow and postdoctoral researcher jointly in the Machine Learning Department at Carnegie Mellon University and in the Booth School of Business at the University of Chicago. Previously, he received his PhD in Computer Science from Purdue University where he was awarded the Bilsland Dissertation Fellowship for his work in the theory of structured prediction. Kevin’s current research interests lie at the intersection of causality and machine learning, and also serves as Production Editor of the Journal of Machine Learning Research.