Weakly Supervised Causal Representation Learning

Explore weakly supervised causal representation learning in this comprehensive talk by Johann Brehmer from Valence Labs. Delve into the challenges of learning high-level causal variables and structures from unstructured, low-level data like pixel-level information. Discover how weak supervision, involving paired samples before and after random interventions, enables learning causal variables and structures. Examine implicit latent causal models and variational autoencoders that represent causal structure without explicit graph optimization. Learn about the application of these models to simple image data for disentangling causal variables and enabling causal reasoning. Gain insights into the current limitations of causal representation learning and potential future developments. The talk covers key topics including structural causal models, theoretical foundations, practical implementations, experimental results, and concludes with an outlook on the field and a discussion session.