This particular feature of artificial intelligence could be harnessed to make it easier for people who cannot speak to communicate.

Another breakthrough in  artificial intelligence  may pave the way for the treatment of pathological conditions. New research, the results of which were pre-published on the arXive platform, showed that artificial intelligence can decode words and phrases by “reading” brain activity.

The training of the algorithm

The researchers first fed an algorithm with 56,000 hours of speech, which came from 53 different languages. This software recognized and recorded the sound quality characteristics of all words heard in speech. Through this process the algorithm essentially “learned” to recognize different words in spoken speech.

At the same time, the scientists invited 169 volunteers to listen to some tape-recorded novels. As the participants listened to the novels, the scientists recorded the activity of their neurons with a neuroimaging technique called magnetoencephalography. In this way, the researchers recorded brain activity while listening to thousands of different words.

Finally, the AI ​​software matched the sound of each word to the brain activity recorded when the participants heard those words in the novels.

Guessing the words

After this training process, the scientists used the algorithm to “guess” which words corresponded to just three seconds of brain activity, as recorded by the method of magnetoencephalography. The algorithm suggested a list of ten words each time, which in 73% of the cases contained the correct word corresponding to the brain activity. As the researchers noted in their relevant statements, this is an important achievement, since the success rate of corresponding methods developed in the past did not exceed 30%.

Important limitations

However, there are still many limitations of this technology. An important limitation is that the algorithm was able to predict these words because it was already trained with a predefined vocabulary. Thus, there can be no practical applications yet, because the algorithm would have to “learn” to recognize millions of words from brain activity. Furthermore, through this particular method, it is extremely difficult to recognize simple words such as “yes” or “no” through brain activity.

The researchers will continue research in this direction, attempting to train the algorithm to more accurately predict words based on brain activity.

Source: 

Τεχνητή νοημοσύνη μεταφράζει την εγκεφαλική δραστηριότητα σε λέξεις

The resouce is in Greek