# Learning to Induce Causal Structure

**Authors:** Nan Rosemary Ke, Silvia Chiappa, Jane Wang, Jorg Bornschein, Theophane Weber, Anirudh Goyal, Matthew Botvinic, Michael Mozer, Danilo Jimenez Rezende  
**arXiv:** [2204.04875](https://arxiv.org/abs/2204.04875)  
**Date:** 2022-04-11  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2204.04875v1/

## Key points

- Paper by Nan Rosemary Ke et al. focuses on inferring underlying graph structures from observational and interventional data
- Novel approach treating the inference process as a black box
- Proposes a neural network architecture for mapping data to graph structures through supervised training on synthetic and naturalistic graphs
- Model shows strong generalization capabilities and outperforms existing approaches on linear continuous and MLP data
- Discusses related works in causal structure learning and potential practical applications of the model

## Summary

The paper "Learning to Induce Causal Structure" by Nan Rosemary Ke et al. addresses the challenge of inferring underlying graph structures from observational and interventional data. Existing algorithms use score-based methods or independence tests to generate candidate graphs, but this work takes a novel approach by treating the inference process as a black box. The authors propose a neural network architecture that learns to map data to graph structures through supervised training on synthetic and naturalistic graphs. Their model demonstrates strong generalization capabilities and outperforms existing approaches on both linear continuous and MLP data. The paper also discusses related works in causal structure learning and highlights the potential for practical applications of their model in realistic settings where accurate causal relationships need to be inferred.