A new brain implant transforms written thoughts into text


“Neural Chain” by Elon Musk already Wave On the technical aspects of neural implants, but it hasn’t shown how we Actual use of implants. So far, the prospect of proving implants is still in the hands of academia.

This week, that community provided a considerable Impressive example The prospect of nerve implants. Using the implant, the paralyzed individual managed to enter about 90 characters per minute just by imagining that he was writing these characters by hand.

Previous attempts to provide typing capabilities to paralyzed people through implants involved providing subjects with a virtual keyboard and letting them use their minds to manipulate the cursor. The process is effective but slow, and requires the user’s full attention, because the subject must track the progress of the cursor and determine when to perform the equivalent key press. It also requires users to spend time learning how to control the system.

However, there are other possible ways to transfer characters from the brain to the page. Somewhere in the thinking process of writing, we form an intention to use a specific character, and using implants to track that intention may work. Unfortunately, the process is not particularly easy to understand.

Downstream of this intent, decisions are transmitted to the motor cortex, where they are translated into actions. Similarly, there is an intent phase where the motor cortex determines that it will form a letter (for example, by typing or writing), and then converts it into the specific muscle movement required to perform the action. These processes are better understood, and they are the goals of the research team for their new work.

Specifically, the researchers placed two implants in the pre-exercise cortex of the paralyzed person. This area is considered to participate in the formation of the intention to perform the action. Capturing these intents is more likely to produce clear signals than capturing the action itself. The action can be complex (any action involves multiple muscles) and depends on the context (your hand relative to the page you are writing, etc.).

Placing the implant in the correct position, the researchers asked the participant to imagine writing letters on the page and recording his neural activity as he did so.

There are approximately 200 electrodes in the participant’s pre-exercise cortex. Not everyone is useful for writing letters. But for those authors, they performed principal component analysis and determined the characteristics that have the greatest difference in neural records when imagining various letters. Converting these records into two-dimensional graphs, it is obvious that the activities seen when writing a single character are always clustered together. And physically similar characters-p with b, For example, or H, n, with [R[RForm clusters close to each other.

(Researchers also asked participants to make punctuation marks, such as commas and question marks, and use> to indicate spaces and tildes over a period of time.)

Overall, the researchers found that they can decrypt the appropriate characters with an accuracy of over 94%, but after recording the neural data, the system requires relatively slow analysis. To make things run in real time, the researchers trained a recurrent neural network to estimate the probability of the signal corresponding to each letter.

Although the amount of data processed is relatively small (only 242 sentence characters), the system works well. The time interval between the idea and the characters appearing on the screen was about half a second. Participants were able to generate about 90 characters per minute, which easily surpassed the previous implant-driven typing records (about 25 characters per minute). The original error rate is about 5%, and the application of a similar automatic correction system may reduce the error rate to 1%.

The tests are all done with prepared sentences. However, once the system is validated, the researchers require participants to enter free-form answers to the questions. Here, the speed dropped slightly (75 characters per minute), and the error rose by 2% after auto-correction, but the system still works normally.


Source link