Plug-and-play (zero-shot) decoding

The aspiration to have a decoder work usefully on a new day, session, or even a new person with no calibration data at all — the user turns the device on and it simply decodes. Two routes lead here. Domain generalization trains on many source sessions or subjects so the learned features are session-invariant and need no adaptation on an unseen target; alignment-based routes require a small slice of unlabeled data to snap the new session onto a reference before a fixed decoder runs, which is calibration-light rather than truly calibration-free.

Genuine zero-shot cross-subject decoding is hardest of all, because different brains and electrode placements do not share a common feature space, and it remains largely aspirational for high-performance intracortical control. It is more mature in EEG, where population-trained models plus a covariance-recentering step can give a workable day-one decoder that a short calibration then improves. Honest reporting separates truly zero-shot performance from performance after a few unlabeled or labeled seconds, because these are very different clinical promises.