Unsupervised decoder adaptation

Adapting a decoder using no behavioral labels at all — not even self-supervised ones inferred from clicks or targets — relying instead purely on the statistical structure of the incoming neural data and priors about the task. Techniques include realigning the input feature distribution to a reference (matching means, covariances, or whole distributions), fitting a latent-variable or hidden-Markov model to the unlabeled neural stream and mapping its recovered states back onto known outputs, and enforcing that decoded outputs obey plausible dynamics (a hand cannot teleport, letters follow a language model).

The appeal is that unlabeled neural data is free and continuous, so adaptation can run silently and forever. The catch follows directly from the shift taxonomy: unsupervised alignment can correct covariate shift but is blind to concept drift and to a global relabeling of outputs (a rotation that swaps left and right can look statistically fine). Practical systems therefore lean on structural priors — output smoothness, language models, task grammars — to break these ambiguities, and many hedge by allowing rare, cheap supervised anchors when confidence drops.