Self-recalibration (semi-supervised adaptation)

Self-recalibration keeps a decoder aligned to non-stationary neural signals without asking the user to stop and perform an explicit calibration task. It obtains the labels adaptation needs indirectly: from context when the goal is externally known (retrospective target estimation), from the decoder's own high-confidence outputs used as pseudo-labels in an expectation-maximization loop, or from distribution alignment that recentres or re-scales incoming features to match a reference without any labels at all — including Riemannian recentring of covariance features and other domain-adaptation techniques.

The appeal is eliminating the daily calibration burden that otherwise gates practical BCI use; the danger is confirmation bias. When pseudo-labels come from the decoder's own predictions, a decoder that has begun to drift will generate labels consistent with its drift and reinforce it, so an unguarded self-recalibrating loop can slowly walk away from good performance. Robust designs mitigate this with confidence thresholds, distribution-shift detection that pauses or resets adaptation, anchoring to a trusted reference model, and label-free alignment methods that do not depend on the decoder being right in the first place.