Reproduce “Learning to Induce Causal Structure” on OpenResearch

About this paper

Nan Rosemary Ke, Silvia Chiappa, Jane Wang, Anirudh Goyal, Jorg Bornschein, Melanie Rey, Theophane Weber, Matthew Botvinic and 2 more

CSIvA learns causal graphs by training an attention-based model on synthetic datasets paired with known graphs, including intervention indicators, then applying the mapping to new systems. On Asia, Sachs, and Child benchmarks, its Dirichlet-trained version reaches Hamming distances of 3, 5, and 10, versus non-CSIvA scores of 4, 9, and 14. These results use simulated benchmark data, not real-world systems.

Key insights

• Interventional samples substantially improve edge recovery, while larger datasets continue to help up to roughly 1,000–1,500 samples; hidden intervention targets and soft interventions cause only modest degradation in the tested settings.
• Sample-level attention is essential: removing it raises Hamming error from approximately 1–6 to about 15–40, whereas removing the auxiliary loss produces a smaller deterioration.
• Synthetic supervision can transfer to simulated Asia, Sachs, and Child Bayesian-network benchmarks, but success depends on the informativeness and diversity of the training distribution rather than eliminating the synthetic-to-real domain gap.

Summary by alphaXiv.