UAI Invited Talk: Learning Causal Variables from Temporal Sequences with Interventions

Abstract

Understanding the causal relations between variables in an environment is a crucial step for intelligent systems. The field of causal representation learning aims at identifying such causal relations, but how can we actually do that? In this talk, I give an introduction to our line of research on using temporal observations, e.g. videos, in which some causal variables have been externally influenced. I show under which situations variables become identifiable, as well as their causal graph. Finally, I conclude with open challenges in causal representation learning.