Machine learning (ML), with its success in language understanding to biological settings, is a key ingredient to intelligent agents that help us with science and decision making. However, ML faces two major hurdles that limit its wider use. First, ML systems struggle to generalize out-of-distribution (OOD), to unseen tasks and domains. Second, ML systems learn correlations but science and decision making require causal inference – an inference about the effects of interventions. The field of causality, with its formalism of causal models, provides a theoretical framework to address the shortcomings of ML systems. Causality benefits from ML too: instead of carefully measuring variables of interest and defining causal models, with ML, we can learn infer quantities from rich sources of data.


In this course, we'll begin with an introduction to the theory of causal models. We'll build on this foundation and study the role causality plays in OOD generalization. Then, we'll study how techniques from ML such as prediction with NNs, representation learning, and gradient based optimization help us leverage large-scale, unstructured data to make causal inferences, from estimating effects to discovering causal models. We'll focus on the challenges and open research problems around learning causal variables and models from data using ML. This is an advanced course, taught seminar-style, and expects students to have a strong background in ML.

The course website from last year can provide a general idea of  how the course will be structured. However,  there will be different readings this year, and the grading scheme will change slightly.

I will assume programming experience and familiarity with topics taught in Fundamentals of Machine learning (or equivalent).  Background in probabilistic graphical models will be useful.