Counterfactual and causal neural modeling

The goal is a model that predicts not just observed neural activity but its response to interventions — what the population would have done under a stimulation pattern, a perturbation, or a different input that was not in fact delivered. This raises the modeling bar from correlation and prediction (Granger-style or encoding models) to Pearl's higher rungs: interventional and counterfactual queries require a causal model of the dynamics, not merely a good fit to observational data. In the BCI setting this is not academic — closed-loop stimulation, cortical microstimulation feedback, and adaptive neuromodulation all pose fundamentally interventional questions.

Two threads define the frontier. One uses perturbations as ground truth: optogenetic, microstimulation, or lesion interventions provide interventional data against which a dynamical model's counterfactual predictions can be validated, turning 'can this model predict the effect of a stimulus it never saw' into a testable claim. The other imports causal representation learning and structural causal models into the latent-dynamics setting, seeking latent factors that are stable and manipulable across interventions rather than merely predictive. Both remain early: purely observational neural data underdetermine causal structure, and validated counterfactual accuracy under novel stimulation is still rare and a meaningful benchmark to demand.