A Foundational Model for Causal Discovery
The Arrow is out and on target:
We have made our paper Arrow: A Foundation Model for Causal Discovery publicly available. Arrow is one of the first foundational models for causal discovery from observational data. Figuring out whether A causes B (not just that they’re linked) usually takes careful, slow analysis. Our new AI model called Arrow skips that: feed it a new dataset, and in one step it predicts a full cause-and-effect map of the variables, guaranteeing the result never contains impossible circular logic.
Arrow was trained on a huge stream of synthetic example datasets covering many kinds of relationships and randomness, and our results show that it matches or beats existing methods, at much lower computational cost.
The potential of this work is huge: Arrow can speed up and improve cause-and-effect research in medicine, economics, education and many other fields significantly.