About the workshop

Machine learning (ML) has established itself as the dominant and most successful paradigm for artificial intelligence (AI). A key strength of ML over earlier (symbolic, logic and rule-based) approaches to AI, is its ability to infer useful features or representations of often very high-dimensional observations in an automated, data-driven way. However, in doing so, it generally only leverages statistical information (e.g., correlations present in a training set) and consequently struggles at tasks such as knowledge transfer, systematic generalization, or planning, which are thought to require higher-order cognition.

Causal inference (CI), on the other hand, is concerned with going beyond the statistical level of description (“seeing”) and instead aims to reason about the effect of interventions or external manipulations to a system (“doing”) as well as about hypothetical counterfactual scenarios (“imagining”). Similar to classic approaches to AI, CI typically assumes that the causal variables of interest (i.e., an appropriate level of description of a given system) are given from the outset. However, real-world data often comprises high-dimensional, low-level observations and is thus usually not structured into such meaningful causal units.

The emerging field of causal representation learning (CRL) aims to combine the strengths of ML and CI. Much like ML went beyond symbolic AI in not requiring that the symbols that algorithms manipulate be given a priori, in CRL low-dimensional, high-level variables along with their causal relations should be learned from raw, unstructured data, leading to representations that support notions such as intervention, reasoning, and planning. In this sense, CRL aligns with the general goal of modern ML to learn meaningful representations of data, where meaningful can also include robust, explainable, or fair.

One aim of this first workshop on CRL is to bring together researchers focusing mainly on either CI or representation learning, from both theoretical and applied perspectives. Moreover, the workshop aims at engaging the various communities interested in learning robust and transferable representations from different perspectives, in order to foster an exchange of ideas. Given that this is still a young, emerging line of research, another goal is to establish a common vocabulary and to identify useful frameworks for addressing CRL.

Selected References

The following is a non-exhaustive list of selected references related to the topics of the workshop, compiled before announcing the call for papers.