Title: Causal Foundations for AI-Driven Decision-Making in Complex Systems
Abstract: How will a small change in atomic structure affect the properties of a material? How will a minor change in financial regulations affect the economy? How will adding one gene to a cell affect its behavior? Increasingly, humanity's progress on pressing scientific and societal problems depends on our ability to predict how different actions (interventions) will affect a complex system, which is intimately related to our *causal* understanding of the system.
In this talk, I will discuss my work on learning causal models of a system from interventional data, covering three main areas of my research: (1) causal structure learning, (2) causal representation learning, and (3) experimental design.
In the area of causal structure learning, I will present an algorithm that addresses the commonly-encountered problem of *unknown* intervention targets, with an emphasis on combinatorial aspects of the algorithm. In the quickly-growing area of causal representation learning, I will present an algorithm that simultaneously learns both (i) a map from high-dimensional data to "macro-variables" that describe the system, and (ii) a causal model over those macro-variables. On the topic of experimental design, I will present one of the first instance-dependent characterizations of experimental complexity, along with an algorithm which matches the instance-dependent quantity up to a logarithmic factor.
I will conclude with an outlook, describing how these areas of research may integrate with other areas to form a foundation for the future of decision-making in complex systems. In particular, I will discuss the importance of analyzing entire "data-to-decision pipelines", along with potential extensions of causal representation learning to other forms of "structure-aligned" representation learning.
Bio: Chandler Squires is a final-year PhD student in Electrical Engineering and Computer Science at MIT, advised by Caroline Uhler and David Sontag.
His research focuses on building the statistical and computational foundations for AI-driven decision-making in scientific applications, especially in the fields of cellular biology and drug discovery. His primary research topics include causal structure learning, experimental design, and causal representation learning. To tackle problems in these fields, his work combines techniques from optimization, statistics, and computer science.
