Open loop vs closed loop
Imagine steering a car with your eyes closed. You can turn the wheel, but with no view of the road you have no idea whether you're drifting left or holding the lane. That's an open-loop system: you act, but you never see the result. A surprising number of early decoders worked this way in the lab — the user imagined a movement, the model made a guess, and nothing came back to the person.
A closed-loop BCI hands the user back their eyes. The decode is turned into something they can perceive — a cursor that moves, a letter that lights up, a robotic hand that reaches — and they use it to correct in real time, just like glancing at the road and easing the wheel back. Suddenly the person isn't a passive signal source; they're a partner steering toward a goal, nudging their own brain–computer interface on course.
Feedback latency
Seeing the result helps only if you see it *soon*. Latency is the gap between your intention and the feedback that follows. When it's short — a fraction of a second — your brain happily ties the two together: "I did that." Stretch the gap, and that feeling of agency frays. By the time the cursor finally moves, you've already tried three other things, and you can no longer tell which thought caused what. The mapping between intention and outcome becomes impossible to learn.
Just as important: the delay must be consistent. A loop that's snappy one moment and sluggish the next is harder to learn from than one that's slow but steady, because your brain can't find a stable rule to lock onto. This is why real-time decoding runs on such a tight budget. Every stage — reading the signal, pulling out features, the model's prediction, drawing the result — eats a few milliseconds, and the whole chain has to finish before the window for "I did that" slams shut.
Neurofeedback: training the brain
So far the loop has pointed outward, toward a cursor or a hand. Neurofeedback turns it inward. Instead of decoding an intention into action, you measure some feature of the person's own brain activity — often the power in a brain rhythm, like how strong a particular oscillation is — and show it back to them directly: a bar that rises, a tone that brightens, a rocket that climbs higher when the target rhythm grows.
Given that mirror, people can gradually learn to nudge the rhythm up or down — not by following any conscious recipe, but the way you'd learn to wiggle your ears or relax a tense shoulder: trial, feedback, and a slow drift toward what works. The brain leans on plasticity, reshaping its own patterns to chase the reward on screen. It's the same closed loop as before, only now the brain is both the driver and the thing being trained.
Shared & assisted control
Even with a fast, well-trained loop, a raw decode is jittery — it twitches and overshoots. The fix isn't always a better model; often it's a better *partner*. In shared control, the machine contributes its own smarts alongside the user: it smooths out the jitter, holds steady when the signal goes quiet, and offers autocomplete the moment your intent is obvious, the way a phone keyboard finishes your word.
This blend is what turns a noisy signal into trustworthy cursor control, or a rough reach into safe prosthetic motion. The assist can also enforce limits the brain shouldn't have to police — a robotic arm that simply refuses to move faster than is safe, or to close hard on a person's hand. The art is balance: too little help and the decode feels unusable; too much and the user feels overridden, no longer the one in charge.