In machine learning, we are often interested in learning predictors and representations of some (high-dimensional) random variable. In contrast, causality allows us to additionally predict how the system underlying the data will respond to external manipulations. This has countless applications, including classic examples like epidemiology or econometrics, but also, more recently, AI for scientific discovery.
In this course, we will cover three fundamentals of causal learning: reasoning, discovery, and representations. More concretely, we will first cover methods to estimate causal effects, interventional distributions, and counterfactuals assuming a known causal structure. Next, we will discuss how the causal structure itself can be learned from data using machine learning techniques. Finally, we will discuss the representation learning setting with high-dimensional measurements, where causal variables themselves are not directly known but have to be learned.
Target group: PhD students in Data Science and Computer Science interested in causality.
Prerequisites: The course assumes no prior background in causality but requires some prior knowledge in statistics and machine learning.
Evaluation: The grade will be determined via assignments (theoretical and practical deepdives covereing a particular lecture/topic), and/or, depending on the number of students, a presentation to the class.
Teaching format: Blackboard
ECTS: 3 Year: 2025
Track segment(s):
Elective
Teacher(s):
Francesco Locatello
Teaching assistant(s):
- Trainer/in: Francesco Locatello