Causal inference with generative neural network models

Research Project, 2025 –

Observational data, unlike data from controlled experiments, contains limited information about causal relationships—illustrated by the common saying “correlation does not imply causation” in the case of two variables. However, when multiple variables interact, observational data does carry some causal information.
Advances in machine learning now allow us to train generative neural network models, which can learn patterns in the data that implicitly reflect these relationships. By using generative neural network models as a stand-in for real data, akin to running simulated controlled experiments on the model itself, we can test causal hypotheses and uncover the causal relationships the model has learned. With causal understanding interventions and further experiments can be designed.