Self-recalibrating decoder

A decoder that refreshes its own parameters during ordinary use, without ever pausing for a cued calibration block, by manufacturing its own training labels from context. The canonical example is retrospective target inference (RTI): in a point-and-click BCI, once the user selects an on-screen item, the system reasons backward that they had been aiming at that item for the preceding seconds, and uses those inferred targets as labels to refit the decoder. Related schemes infer intent from click confirmations, from the assumption that motion between selectable items is goal-directed, or from stereotyped task structure.

This is best understood as self-supervised rather than truly unsupervised: the labels are not hand-annotated, but they are extracted from task structure and user actions that stand in for ground truth. It has been demonstrated to sustain intracortical point-and-click performance in people with tetraplegia across many months with little or no explicit recalibration. The obvious failure mode is when the inferred labels are wrong — free-form or exploratory use, or an interface with no discrete selection events, gives the method nothing reliable to learn from.