Closed-loop decoder adaptation (CLDA)

Closed-loop decoder adaptation is the family of methods that update the decoder's parameters while the user is actively controlling the effector, using the ongoing closed-loop data rather than a separate offline calibration block. Because the decoder changes based on data generated under its own current behavior, and the user simultaneously adapts to the decoder, CLDA is explicitly a solution to the two-learner problem: it steers the machine side of the co-adaptation deliberately rather than leaving it fixed. Concrete schemes differ in what they update (tuning/observation model, dynamics, or both) and how fast (per-trial, per-batch, or smoothed).

A representative approach fits an initial (often assistive or seeded) decoder, then repeatedly re-estimates its parameters from short recent windows of closed-loop activity paired with intention-inferred kinematics, blending each new estimate with the previous parameters by a smoothing factor to avoid instability. This yields decoders that come online quickly with little calibration data and that track slow changes in the neural signal. ReFIT is a two-stage special case of the same principle; SmoothBatch and adaptive Kalman methods generalize it to continual updating.