Neurofeedback · Lobeworks/17

Neurofeedback shows a person, in real time, a measure of their own brain activity and rewards them when it moves the way it should, and it is how a brain learns to produce a signal on purpose: the basis of training a pattern for therapy, and of the half of every brain-computer interface in which the user, and not the d

Neurofeedback. Neurofeedback shows a person, in real time, a measure of their own brain activity and rewards them when it moves the way it should, and it is how a brain learns to produce a signal on purpose: the basis of training a pattern for therapy, and of the half of every brain-computer interface in which the user, and not the decoder, does the learning.

The loop has four parts: record a feature of the activity, turn it into something the person can perceive (a tone, a bar, a cursor), reward the change wanted, and repeat. In 1969 Eberhard Fetz showed that monkeys rewarded for raising the firing of single cortical neurons, with a click or a light as feedback, learned to do it, which was the first evidence that the activity an interface reads is not fixed and can be trained.

What can be learned has a shape. Sadtler and colleagues changed the map from neurons to a cursor and found that, within hours, monkeys learned maps that kept to the patterns their neural population already produced, and struggled with maps that needed patterns outside them. Fast learning stays inside what a network already does.

It makes the user a second learner. A decoder that adapts while the person adapts can converge on a skill or chase it, and the success of the pair depends on both rates.

The effect needs controls to be believed. Feeling better after sessions of feedback can come from the attention, the expectation or the practice, so a therapeutic claim needs a comparison with sham feedback that looks the same.

Questions: Who learns in a brain-computer interface, the decoder or the brain? Both. The decoder learns to map the activity it records onto what the person intends, and the person, watching the result, learns to produce activity the decoder reads well, as monkeys learned to raise the firing of single neurons for a reward in 1969. The human side learns fastest when the new map asks for patterns its neurons already produce. Because both adapt at once, an update to the decoder changes a skill the person built around the old one, which is why a recalibration is closer to a change in treatment than to a software patch.