How do researchers find causal neurons?

Lens adapted to this topic: Limits and competing interpretations

A skeptical interpretation emphasizes that no single test automatically identifies a uniquely causal neuron. A highly active neuron may be a bystander, while redundant neurons may each appear dispensable. Network transformations, hidden variables, finite samples, and violations of causal-discovery assumptions can produce misleading graphs or perturbation results. More exhaustive interventions and explicit counterfactual analysis are therefore needed.

• Lesioning one element at a time can give misleading estimates when causal contributions depend on combinations of elements; broader perturbation regimes are recommended.
• Activity patterns may be poor indicators of causal contribution because network transformations can dissociate activity from functional influence.
• Neural causal-discovery methods may fail to recover the true graph in finite samples, particularly when the faithfulness assumption is violated.
• Counterfactual intervention tests can miss causes that are independently sufficient and can be difficult to interpret because neural dependencies are not generally transitive.
• Mediation—the role of neurons in the causal chain from stimulus to behavior—is conceptually important but difficult to measure with conventional tools.

Featured sourceDownstream network transformations dissociate neural activity from causal functional contributions | Scientific Reportsnature.com6 sourcesView sources