Advances in Differentiable DAG Learning: Analytic Constraints and Truncated Power Iteration
This project integrates two recent advances in differentiable DAG learning: (1) analytic DAG constraints for improved computational stability and (2) Truncated Matrix Power Iteration (TMPI) for escaping gradient vanishing. The methods are applied to causal discovery and structure learning in high-dimensional settings.
Oct 1, 2021
