An AI developed by Facebook鈥檚 owner Meta can scan a person鈥檚 brainwaves to 鈥渉ear鈥� what someone else is saying to them. The software is of limited practical use currently, but could be a stepping stone to reading a nonverbal person鈥檚 thoughts to allow them to better communicate 鈥� though Meta has no immediate plans to turn it into a commercial product.
at Meta and his colleagues used existing data sets of brainwaves taken from 169 people as they listened to recordings of people speaking. The team split the data into chunks of 3 seconds and fed both the brainwaves and the sound files to an AI, which learned to spot patterns between the two.
The researchers held back 10 per cent of the data for testing, giving the AI brainwave data that it had never had before. They found that it was able to deduce which individual words, from a vocabulary of 793, the person was listening to at the time 鈥� at least to an extent. For each recording, the AI predicted a list of 10 words, and 73 per cent of the time this list included the correct word.
Advertisement
at Imperial College London says that more refinement would be needed before such a system was accurate enough to be useful. He also questions whether the type of brain scans the team used 鈥� magnetoencephalography and electroencephalography, which are both non-invasive 鈥� could ever provide enough detail of the brain鈥檚 state to increase accuracy.
鈥淚t鈥檚 about information flow,鈥� he says. 鈥淚t鈥檚 like trying to stream an HD movie over old-fashioned analogue telephone modems. Even in ideal conditions, with someone sitting in a dark room with headphones on, just listening, there are other things going on in the brain. In the real world it becomes totally impossible.鈥�
King says that the research goes one step further than similar previous work training an AI to identify which word a person was thinking of, or song they were listening to, from a small group of options. He says that determining what words a person is hearing is of limited use, and that the next step for his team is to try to interpret a person鈥檚 thoughts and allow them to communicate.
鈥淚 think the next step is going to be language production. We鈥檙e not decoding thoughts at this stage,鈥� he says. 鈥淚 mean, we don鈥檛 have any clue whether this is possible or not. But given how hard it is to record even something as crisp as the sound that you hear, language production, I think it鈥檚 going to be even harder.鈥�
King says that the work was carried out by the company鈥檚 scientific arm, Meta AI, which publicly releases general machine-learning research for use by other researchers and has no obligation to create practical applications. 鈥淲e鈥檙e not trying to build any product at this stage,鈥� he says. 鈥淣ow, whether there will be applications of that research, I think it鈥檚 possible, but it鈥檚 also a bit more distant.鈥�
Reference:
