Scientists went one step further to create machines that understand what a human being is saying, analyzing only his brain cells and then automatically converting his sentences into written text.
Researchers in the US have developed an artificial intelligence algorithm that can decode the brain’s neuronal activity and translate it in real-time into highly accurate sentences (only 3% error), which is achieved for the first time at this level.
Brain-machine interfaces have so far had limited success in decoding thought-based brain activity alone and thus producing artificial speech, achieving precision far less than physical speech, and have so far failed to do so; to read whole sentences rather than just words.
This time, the researchers, led by Joseph Makin and Edward Chang of the University of California-San Francisco, who published this in the journal “Nature Neuroscience”, took advantage of the latest developments in machine translation field to train artificial neural networks to directly convert neuron signals into sentences.
Four volunteers, whose skulls had been implanted with neuronal activity recording electrodes, were reading out loud text sentences. The artificial intelligence system thus learned to correlate neuronal signals with speech and word components (vowels, etc.), and then learned to “bind” words into sentences.
The accuracy of the algorithm is analogous to that used by “smart” business computer systems of automatic conversion of speech into written text.
However, according to researchers, the system still needs further improvement, as it has, so far, been unable to exceed 30 to 50 sentences, based on a vocabulary of up to 250 words.
The next step will be to enable the system to convert the neuronal activity of people who are just thinking and not talking into sentences.