KAUST Rising Stars in AI: Causal Representation Learning

Abstract

I gave an invited talk at the Rising Stars in AI Symposium 2023 at KAUST! I covered our recent works on Causal Representation Learning, why this task is important, and what its current state is. The discussed papers are linked above, and my abstract below:

Future AI systems will need to reason about their environments in order to be accurate, safe, and robust. One promising direction of achieving this is via causal reasoning, e.g., by exploiting the knowledge of cause-effect relations in an environment. In practice, however, we commonly do not know the causal variables and their relations up-front, but instead rely on much higher-dimensional observations like images. Hence, in this talk, I will give an introduction to causal representation learning as a tool for identifying the causal variables in such environments. I will focus on our recent line of work on Causal Identifiability from Temporal Intervened Sequences (CITRIS), arguing that the causal variables one finds depend on the actions/interventions available. Further, we discuss the open challenges and possible future directions of causal representation learning.