Semi-supervised Learning of Visual Causal Macrovariables

Discovery of causally related concepts is one of the key challenges in extracting knowledge from observational data. Lower-dimensional “causal macrovariables” represent concepts which preserve all relevant causal information in high-dimensional systems. Existing causal macrovariable discovery algorithms are limited by assumptions about known and controllable interventions. We propose a variational autoencoder-inspired architecture with regularization terms for semi-supervised causal macrovariable discovery. These terms impose domain knowledge regarding visual causal concepts to differentiate between correlation and causation. Experiments on both synthetic and real-world datasets with known causal dynamics show that our method can discover more concise and precise causal macrovariables than unsupervised methods.