By Shriyadita De
Imagine yourself sitting in a seminar room with one of the world’s most influential theoretical physicists, Dr. Stephen Hawking, giving a lecture on his discoveries of black hole radiation and the origin of the universe. Due to his paralysis from motor neurone disease, he was unable to communicate fluently and was restricted to about 15 words per minute. You are astonished to hear his lecture, but would it not have been more advantageous if Dr. Hawking had a device that would facilitate communication with others more freely?
The Idea
Using a novel AI model developed by Meta AI, scientists are getting closer to decoding words and sentences from brain activity. The model has the correct answer among its top 10 predictions 73% of the time, after only 3 seconds of brain activity. It can be used to help thousands of people around the world who have difficulty communicating verbally or physically.
Converting brain activity to speech has been an everlasting challenge for scientists, and this technology is helping to overcome it. Developing a technology that does not require brain surgery was one of the biggest challenges. Existing technologies require dangerous surgery and the implantation of electrodes. A technology that avoids these invasive procedures would make it more accessible and safer for the general public. To train the AI model, Meta AI scanned 169 volunteers’ brains using electroencephalography (EEG) and magnetoencephalography (MEG) while listening to audiobooks and individual sentences to measure the magnetic or electrical fields of the brain.
The Efficiency
After training the model with these measurements, they looked at brain activity and predicted what the person might have heard within 3 seconds, giving more than 1000 possibilities. Meta AI found that the correct answer was in the first ten guesses, 73 percent of the time when using MEG, making it more effective than EEG.
The model is remarkable, but it needs to be improved. “In this work, we focused on decoding speech perception, but the ultimate goal of enabling patient communication will require extending this work to speech production,” says Meta AI research scientist Jean Remi King. The current model cannot be put to practical use as of now, as it requires a big, expensive machine, which would not make it easily accessible to the public. Moreover, the AI was trained on the limited number of correct words the user may be thinking about. “With language, that’s not going to cut it if we want to scale to practical use, because language is infinite,” says linguist Jonathan Brennan at the University of Michigan.
Last Thoughts
Although the model is not yet implementable, it is the beginning of neuroscientists tackling this challenge. A first prototype is never flawless. With technology and the future in our hands, we can bridge this communication gap and victory will be ours. So, the next time you wish you could read someone else’s mind, think about the challenges scientists are facing right now to make it happen, and what we can achieve in just a couple of years.
Written in 2022
Resources:
How Intel Gave Stephen Hawking a Voice – Wired
An AI can decode speech from brain activity with surprising accuracy
Meta AI researchers develop ways to read speech from people’s brains
BrainBox 
Leave a comment