Framing it as distribution shift
The clean way to think about all of this: training and deployment draw from different distributions. In the language of machine learning it is covariate shift and, more painfully, a drift in the class-conditional encoding itself.
Neural non-stationarity is a drift in the encoding p_t(x|y). A fixed decoder implicitly assumes the day-0 posterior p_0(y|x); once the encoding moves, that posterior is mis-specified.
Note which way the shift runs. In computer vision, covariate shift usually means the inputs p(x) move while the labelling rule p(y\mid x) is fixed. In BCI the encoding p_t(\mathbf x\mid y) moves — the same intent produces different neural features over time. That is why generic domain-adaptation recipes need neural-specific structure to work here at all.
Intention-based relabelling: ReFIT and CLDA
The breakthrough that made intracortical cursor control robust was realising you often know the label without asking. If the user is trying to hit a visible target, their intended velocity at each instant points at that target. ReFIT exploits exactly this.
ReFIT's intention estimate: rotate the observed cursor velocity to point straight at the known goal while keeping its speed. This manufactures self-supervised labels from task structure, with no extra calibration block.
With those inferred labels you refit the Kalman filter — the 'recalibrated feedback intention-trained' step that gives ReFIT its name. Closed-loop decoder adaptation (CLDA) generalises the idea: update decoder parameters during use from continuously inferred intent, so the decoder tracks slow change instead of waiting for the next calibration block.
Co-adaptation: two learners, one loop
Once the decoder adapts online while the user is also learning, you have two adaptive systems inside one feedback loop — the two-learner problem. Done well, co-adaptation is a virtuous circle: the user shapes stable, decodable strategies while the decoder meets them halfway. Done badly, the two chase each other into instability.
The ceiling of supervised recalibration
Why not stop here? Because supervised, closed-loop recalibration has a ceiling made of three limits. First, the recalibration burden never fully disappears — CLDA reduces but does not eliminate the need for structured tasks and occasional ground truth. Second, it assumes a decodable label source; much of what we ultimately want to decode — mood, memory, free cognition — has no visible target to infer intent from. Third, it tracks slow drift well but copes poorly with abrupt change: a loosened array connector, an overnight jump in the unit population.
This is the honest handoff to the frontier. Closed-loop supervised methods are mature, deployed, and genuinely good at slow drift with a known task. Everything beyond that — labels-free, task-free, robust to abrupt change — is where the research of the last several years lives, and where the rest of this track goes.