Generalization to unseen behavioral states

Characterize how well the MAML-pretrained temporal basis function model generalizes to behavioral states that are absent from its calibration data, including clinically relevant states such as rest, movement, sleep, and emotional arousal.

Background

The evaluation trains models on the early portion of each session and tests them on the late portion, but it does not intentionally vary behavioral state between calibration and deployment. Because neural stimulation responses may differ across behavioral and clinical states, the authors identify uncertainty about whether the model can handle state-level distribution shifts that are not represented during calibration.

References

While known behavioral states may be captured as additional covariates that can be input to our basis weight generator, it remains unclear how well our model will generalize to unseen behavioral states.

— Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining  (2608.26649 - Bryan et al., 27 Aug 2026) in Section Discussion, subsection “Limitations,” subsubsection “Evaluation under distribution shift due to behavioral states”