Causal Representation Learning: A Quick Survey

Causal representation learning (CRL) has recently become an object of intensive research. Representation learning aims to infer a lower dimensional, but meaningful, representation from a given set of data, effectively increasing interpretability and processability. Disentangled representation learning applies an independency constraint onto the inferred latent variables of the representation.
The applicability of such frameworks on real world data is limited, as absolute independence between all generating factors is rarely the case. CRL assumes causal relations between these latent factors making it more flexible and suitable for real world settings. In this work we give an overview over several approaches to disentangled representation learning and give a short introduction to variational auto-encoders and generative adversarial networks. We follow up by covering the current state-of-the-art CRL frameworks and finish with remarks regarding current weaknesses of CRL as well as potential research topics.