Advancements in Causal Abstraction: Definitions and Learning Methods
Riccardo Massidda (Ph.D. Student)
In several applications, describing a system at different levels of detail and granularity enables faster reasoning on the model and eases its interpretability. In the context of graphical causal models, Causal Abstraction is a recently defined framework that enables concise representations of low-level systems with numerous variables through smaller high-level models retaining the causal properties such as interventional or counterfactual distributions. Despite preliminary applications in Neural Network interpretability and potential uses in large systems such as weather or brain activation patterns, the existing literature either focuses on discrete models or does not cover the problem of learning causal abstractions from data. The PhD project tackles these two main issues, encompassing both theoretical aspects of the definition of Causal Abstraction and the retrieval of abstract causal models from data. First, on the theoretical side, the project aims to study the problem of abstracting soft interventions and the necessary and sufficient conditions for abstracting different model classes, such as linear models, additive noise models, or post-nonlinear causal models. Then, the project formalizes the problem of Causal Abstraction Learning and provides identifiability results to recover the abstraction function from data depending on the previously mentioned classes and the observability of abstract information such as observational or interventional samples or prior knowledge of abstract variables. Finally, the project proposes practical methods to learn abstraction functions and abstract models from data by combining existing causal discovery approaches to handle data with different granularity.
| Primary Host: | Davide Bacciu (University of Pisa) |
| Exchange Host: | Sara Magliacane (University of Amsterdam & MIT-IBM Watson AI Lab) |
| PhD Duration: | 01 October 2021 - 31 January 2025 |
| Exchange Duration: | 01 September 2023 - 29 February 2024 - Ongoing |
