Causality and Machine Learning

Catalog description

In the past decades, significant progress has been made in tackling long-standing causality problems, such as discovering causality from observational data and inferring causal effects. Moreover, it has recently been shown that the causal perspective aids in understanding and solving various machine learning problems such as transfer learning, out-of-distribution prediction, disentanglement, representation learning, and adversarial vulnerability. Accordingly, this course is concerned with understanding causality, learning it from observational data, and using it to tackle other learning problems. The course covers representations of causal models, how causality is different from association, methods for causal discovery and causal representation learning, and how causality enhances advanced learning tasks, including generative AI. We will address the following questions. Why is causality essential? How can we learn it, including latent variables, from observational data? What role does causality play in learning under data heterogeneity? Can causal principles make generative AI more controllable and capable of extrapolation? How can deep learning benefit from a causal perspective? Two main causality problems are emphasized. One is causal discovery or causal representation learning. It is well known that "correlation does not imply causality," but we will make it more precise by asking what assumptions, what information in the data, and what procedures enable us to successfully recover causal information. Causal relations may happen among the underlying hidden variables—we will also see how to uncover the underlying hidden "causal" variables as well as their causal relations from the measured variables. The other is how to properly make use of causal information. This includes identification of causal effects, counterfactual reasoning, and improving machine learning…