Closed-loop decoder adaptation (CLDA)

The established, largely supervised way of keeping a decoder current while the user is actively using it: update decoder parameters online from the ongoing stream of neural data plus an estimate of what the user intended. In the ReFIT approach (recalibrated feedback intention-trained Kalman filter), intention during a reaching task is estimated by assuming the user always meant to point straight at the target and rotating the training velocities accordingly; SmoothBatch and related CLDA schemes blend old and new parameters gradually to avoid destabilizing the loop. These methods, developed in the graduate-era of intracortical BCI, substantially improved control and are the baseline the frontier now builds on.

Their limitation, and the reason drift-robust research pushes past them, is the supervision they assume. Intention estimation works cleanly in cued, structured tasks (reach to this target) but not during free, unstructured daily use where the true intent is unknown. The frontier goal is to keep CLDA's continual-update spirit while replacing its supervised intention signal with an unsupervised, self-supervised, or context-inferred one, so adaptation can continue during ordinary use with no cued blocks.