A new brain implant turns the thoughts of writing into text
Elon Musk’s Neuralink has been making waves In terms of neuronal implant technology, but it still hasn’t shown how we could actually use implants. For now, showing the promise of implants remains in the hands of the academic community.
This week, that community has pretty much given up an impressive example Which prescribes neuronal implants. Using the implant, a stopped person was able to write about 90 characters per minute imagining that he was writing these characters by hand.
Previous attempts to give paralyzed people the ability to write through implants have allowed them to give the subjects a virtual keyboard and maneuver the cursor with themselves. The process is efficient but slow and requires the full attention of the user, as the subject must monitor the progress of the cursor and determine when to do the equivalent of pressing the key. In addition, the user must spend time learning to control the system.
But there are other ways to get the characters out of their brains and get to the page. Somewhere in the thought process of writing, we create the intention to use a specific character and using an implant to track that intention can work. Unfortunately, the process is not particularly well understood.
Down the river of that intention, the decision is transmitted to the motor cortex, where it becomes action. Again, there is an intentional stage in which the motor cortex determines that it will form the letter (in writing or in writing, for example) and then return to the specific muscle movements required to perform the action. These processes are much better understood, and the research team is focused on their new work.
Specifically, the researchers placed two implants in the premotor cortex of a disabled person. This area is believed to be involved in shaping intentions to make movements. Catching these intentions is much more likely to create a clear signal than catching the movements themselves, as they are likely to be complex (any movement involves multiple muscles) and depends on the context (your hand is on the side of the page you are writing, etc.).
When the implant was in the right place, the researcher asked the participant to imagine writing the letters on a page and recorded the neural activity as he did so.
In total, there were approximately 200 electrodes in the participant’s premotor cortex. All of them were not for writing information letters. When this was the case, the authors conducted an analysis of the main components that identified the characteristics of the most differentiated neural recordings when several letters were imagined. Turning these recordings into two-dimensional plots, it was obvious that the activity that was seen when writing a single character was always included. And physically similar characters—or and b, for example, or h, n, and r—The formed clusters are close to each other.
(The researchers asked the participant to make punctuation marks such as a comma and a question mark, and> used one to indicate a space and a tilde temporarily).
Overall, the researchers found that the right character could be deciphered with more than 94 percent accuracy, but the system required relatively slow analysis after recording neuronal data. To make things work in real time, the researchers trained a repetitive neural network to calculate the probability of a signal corresponding to each letter.
Despite working with a relatively small amount of data (only 242-character characters), the system worked very well. The difference between the thought shown on the screen and a character was about half a second, and the participant was able to produce about 90 characters per minute, surpassing the previous record for implant-guided writing, which is about 25 characters per minute. The gross error rate was about 5 percent, and applying a system that can automatically correct typing can lower the error rate to 1 percent.
The tests were performed with prepared sentences. After validating the system, however, the researchers asked the participant to write free answers to the questions. Here, the speed dropped slightly (75 characters per minute) and errors increased to 2 percent after self-correction, but the system continued to work.