Emergent representations of graphical structure in mechanistic neural models of causal judgment

Task-optimized recurrent neural networks learn causal judgment from stochastic samples and develop neural activity with an emergent geometry that reflects the underlying causal structure.

• Why it matters: Prior causal cognition work emphasizes normative behavior, but it remains unclear how biologically plausible neural systems could learn causal structure from noisy, limited observations and update representations with new evidence.
• What they did: The authors trained recurrent neural networks on a query-based task using unknown causal structures among variables like A, B, and C, then analyzed trained dynamics to extract low-level neural mechanisms.
• The result: The networks achieve high accuracy, use information from non-queried variables in their judgments, and dynamical systems analysis identifies mechanisms that represent causal relationships, enabling a bottom-up route for studying causal cognition in the brain.

AI summary of the abstract · may contain errors · not medical advice · read the paper.