Unlocking Generalization Capabilities for Causal Graph Structure Induction with DeepMind, Mila & Google Brain

DeepMind, Mila & Google brain Enable Generalization for Causal Graph Induction

Finding the causal structure and relationships of a system is a difficult problem that affects many scientific disciplines, from medicine to biology and economics. Researchers typically use the graphical formalism known as causal Bayesian network (CBNs), to generate a graph that best represents these relationships. However, unsupervised scores-based approaches to this problem can lead to a prohibitively high computation burden.

The new paper Learning to induce Causal Structure by a research team from DeepMind and Mila-University of Montreal, and Google Brain, challenges the conventional causal approach. They propose a neural network that can learn the graph structure of observational or interventional data through supervised training using synthetic graphs. The proposed Causal structure Induction via attention (CSIvA), developed by the team, effectively turns causal induction into a black box problem. It also generalizes well to new synthetic graphs and naturalistic graphs.

The team summarises their main contributions in:

Source:

DeepMind, Mila & Google Brain Enable Generalization Capabilities for Causal Graph Structure Induction