by Leon Sierau, Institute for Data Science Foundations - on Thursday, October 22, 2026 at 02:30 pm - in room 5.002, HIPone, Blohmstr. 15, 21079 Hamburg.
We will discuss a formalism called “causal abstraction” and how it can be helpful in learning a causally meaningful representation of some data generating process. My proposal is that the goal of most approaches to causal representation learning, namely finding a “disentangled” representation of the true data generating process, is highly unrealistic in most scenarios. Among other reasons, I believe this to be the case due to the restrictive assumptions one has to impose in order to identify a disentangled representation (most critically including injectivity of the observation mechanism and equality of the learners internal state space’s dimension to that of the data generating “outer world”). Causal abstraction allows us to formulate a connection between the “outer world” and a lower dimensional representation that preserves relevant parts of the causal structure. I will show that in situations where the identification of a “disentangled” representation of the data generating process is impossible, one can still find a valid causal abstraction of it. If time allows it, we might also cover some practical approaches on how to construct a causal abstraction from observed data.
My notion of causal abstraction is based on the seminal works of Beckers & Halpern (2019) and Rubenstein et al. (2017), in case you would like to familiarise yourself with it beforehand. However, a systematic discussion of their approaches and how my own formulation is inspired by them is not the focus of the talk.
