A1636
Title: Extracting causal variables through causal representation learning
Authors: Hiroshi Morioka - Shiga University (Japan) [presenting]
Abstract: Causal discovery aims to estimate causal relationships among variables, typically under the assumption that the relevant causal variables are directly observed. In practice, however, it is often unclear what the true causal variables are. This challenge motivates causal representation learning (CRL), which seeks to recover both latent causal variables and the causal relations among them from complex observations. The CRL perspective reveals fundamental limitations of this problem: it is generally ill-posed and cannot be solved without suitable inductive biases on the underlying generative process. A recent CRL framework demonstrates that identifiability can be achieved under assumptions based on multimodal observations. These results illustrate how structural assumptions on data generation make it possible to recover meaningful causal variables from high-dimensional data, extending causal discovery beyond the classical setting of directly observed variables.
