They learned artificial intelligence to read brain activity
A group of American scientists have created a neural network that can read the activity of the cerebral cortex during speech, convert received signals into code, and reproduce what one says. In the future, scientists say, this can help dumb people speak.
American scientists Joseph Mike, David Moses, and Edward Chang learned artificial intelligence (IE) to read brain activity and understand what a person says in this moment. The results of their research have been published in the journal Nature Neuroscience.
So far, the neural network model created by the group of scientists is only working on a limited amount of initial data with predetermined brain activity, which is responsible for human speech. However, they were able to achieve the accuracy of reproduction of what was said at 97%, reports FOCUS.
As Dr. Christian Herf of the University of Maastricht notes, ” have reached a level of accuracy that no one has been able to achieve so far. ”
Researchers themselves say that the IE they created can become an innovative technology that will enable dumb people to talking. “We are still a long way from that, but we believe this technology can become the basis for a speech prosthesis,” wrote The Guardian, quoting Makin.
To create artificial intelligence that can reading the thoughts, the scientists asked several volunteers to read sets of short sentences briefly. In this case, only the phrases, which were repeated by at least three different participants, were selected for the neural network itself.
As a result, scientists have collected around 50 phrases consisting of 250 unique words. Electrodes were placed during the reading of the participants in the experiment, which made it possible to read their brain activity. The received data was loaded into a program that turned them into sets of numbers.
Then the neural network went into action, trying to reproduce the originally read text from the received digits. At first attempts, the system emitted unrelated sounds. But it contained original phrases that she used to compare her own text each time. At the same time, the system emphasized the logic of constructing phrases. As a result of many attempts, scientists have been able to achieve unprecedented accuracy of reproduction with errors of only 3%.
It is true that the scientists themselves recognize that despite the successful result, the implementation of this technology is still far .
After all, their neural network learns from a very small amount of data, and as their volume increases, the error rate increases dramatically. However, if only one person talks about the experiment, the system learns faster. This way she can become an ideal helper for dumb people in the future.