Test-time adaptation

Adapting a model to each new batch of data at the moment of inference, with no labels, using only the unlabeled test data itself. Borrowed from mainstream machine learning, the canonical moves are: recomputing normalization statistics from the current data (test-time batch-norm adaptation), minimizing the entropy of the model's own predictions so it commits confidently to some output (as in the TENT method), and test-time training, where a self-supervised auxiliary loss is optimized on the incoming data to nudge the features back into distribution before decoding.

For BCI, test-time adaptation is attractive because non-stationarity arrives continuously and the test data is exactly where the shift shows up. It is lightweight and needs no stored calibration set. Its risks are specific to closed-loop use: entropy minimization can collapse to a confidently-wrong fixed point, and adapting on a running stream can chase transient artifacts. Because a BCI acts on its outputs in real time, an unstable test-time update can drive the user into a feedback loop that the same update then reinforces, so conservative update rates and safety fallbacks matter more here than in offline benchmarks.