A brain-computer interface has a problem that doesn’t appear in the demonstrations: the brain changes. The electrical signals produced by neurons are not exactly the same every day and neither are those recorded by the electrodes. That’s why, a system that learns to interpret a person’s intent may need periodic adjustments to continue doing solike a clock that updates depending on the time zone we are in.
Now Neuralink, Elon Musk’s company specialized in brain implants, claims to have found a way to reduce that problem. Participants in their clinical trials have used their implants for more than 50,000 hours, accumulating one of the largest data sets of intracortical neuronal activity ever assembled. And it is precisely these records that have been used to pre-train an artificial intelligence model intended for brain-computer interfaces.
The idea is simple. Until now, the system required calibration sessions in which it learned what patterns of brain activity corresponded to certain actions, such as moving a cursor. For this Neuralink has used thousands of hours of unlabeled brain activity for a model to first learn what the usual patterns of neural signals are. That knowledge can then help the decoder interpret the user’s intentions even if the signals change over time. The model uses self-supervised learning to transform the electrical signals associated with the activity of neurons into more stable representations. Thanks to it some decoders can remain operational for weeks rather than days and that, for some participants, calibration has gone from about ten minutes a day to about ten minutes a week.
But the announcement includes another figure that is more difficult to interpret: 11.32 bits per second. It is not a speed. The cursor does not move 11.32 pixels or 11.32 centimeters every second. In a brain-computer interface, bits per second are a measure of the amount of information the user manages to transmit to the computer during a given task. To calculate it, it is taken into account, among other things, the speed and precision with which it manages to select objectives.
Let’s imagine a screen full of buttons. If a person can quickly indicate which one they want to select and barely make mistakes, they are transmitting more information per second than if they take a long time to find it or make mistakes frequently. Bits/s condense that performance into a number. So Neuralink’s 11.32 bits/s says something about the controllability of the interface, not how fast the cursor physically moves. The participant identified as P15 reached 11.32 bits/s, according to Neuralink, which, in simple terms, It would be equivalent to correctly choosing a color among about 2,500 possibilities. Musk’s company claims that the new figure exceeds the performance of most able-bodied people, although that comparison comes from Neuralink and has not been independently verified.
There is another result that is perhaps more important for everyday use. Neuralink has shown the so-called “teleported cursors”: instead of moving the pointer across the screen, the system tries to directly identify the target that the person wants to select and takes the cursor there. It is a demonstration of where the development of these interfaces is heading: that the computer interprets the intention instead of requiring the user to control every small movement.
It is worth putting one last figure in context. The 50,000 hours are data accumulated during actual use of the implants by trial participants; The 11.32 bits/s, on the other hand, correspond to the result of a specific cursor control test. They are not two equivalent measurements nor do they represent the same thing. And the results have been reported by Neuralink in a technical update, not an independent peer-reviewed study.
But together they tell an interesting story. For years, brain-computer interfaces have attempted to teach a machine to translate neural activity into commands for a computer. Neuralink is now trying to use thousands of hours of operation so that the machine itself learns how that language changes over time. The ultimate goal may not be to get a person to move a cursor with their thoughts. That is now possible. Neuralink’s immediate goal is to make a brain-computer interface more stable and universal. But if its founding model manages to learn neural patterns on a much larger scale, the project could become something more ambitious and coherent with the intentions of Elon Musk regarding artificial intelligence: he could be teaching an AI how the brain speaks.