Covariate Shift & Nonstationarity

Neural feature distributions drift within and across sessions from electrode movement, impedance change, fatigue, attention, and altered mental strategy, so the training (calibration) distribution differs from the test (online, later-session) distribution. When the feature distribution p(x) shifts but the labeling rule p(y|x) is stable it is covariate shift; when the mapping itself changes it is concept drift. Either way, a decoder validated only within its calibration session overestimates deployed performance.

This is both a validation concern — honest evaluation requires cross-session or online testing, never a within-session random split alone — and a modeling target addressed by adaptive or recalibrating decoders, unsupervised feature alignment (for example Riemannian recentering, or whitening to a running reference), and stationary feature design. Reporting within-session accuracy as if it predicted daily use is one of the field's most consequential over-optimisms.