Reproduce “Causal Abstraction Inference under Lossy Representations” on OpenResearch

About this paper

Kevin Xia, Elias Bareinboim

Researchers from Columbia University's CausalAI Lab introduce "projected abstractions" to enable formal causal reasoning with abstract, lossy representations, overcoming limitations of the Abstract Invariance Condition. Their framework includes an algorithm for constructing consistent high-level causal models and a new "partially projected C-DAG" for inference, validated by improved causal effect estimation and high-quality generative sampling from binary representations on Colored MNIST.

Key insights

• Causal abstraction can be formally achieved even with lossy representations by systematically incorporating lost information into the high-level model as exogenous variables.
• Standard graphical models are insufficient for causal inference in lossy abstract spaces, necessitating a new graph structure (partially projected C-DAG) to correctly capture dependencies.
• Identifying a low-level causal query within the projected abstraction framework is equivalent to solving the classical identification problem on its high-level counterpart using the partially projected C-DAG, simplifying inference.

Summary by alphaXiv.