Scientists created neurochip that turns thoughts into words with accuracy of over 90%

By Alimat Aliyeva

Swiss scientists from the Federal Polytechnic University of Lausanne (EPFL) have developed a miniature brain-computer interface that converts neural signals into text with high accuracy and low power consumption, Azernews reports.

The chip uses a new approach based on the identification of specific neural markers for each letter, which reduces the amount of processed data to one hundred bytes per marker. This makes the device more compact and less invasive compared to analogues such as Neuralink.

The chip decodes brain activity and converts it into text with 91% accuracy.

The device, dubbed the "miniature brain-computer interface" (MiBMI), is extremely small: it consists of two thin chips measuring only 8 mm2. For comparison, Elon Musk's Neuralink neuroimplant is much larger — about 23×8 mm. The EPFL chip consumes little energy, is considered minimally invasive, and is a fully integrated system that processes data in real time. Neuralink, in turn, requires the installation of 64 electrodes in the brain and processes the data through an application on an external device.
Like other brain-computer interfaces, the new chip mainly tracks the electrical activity of the brain and, based on data from previous neurobiological studies, converts this activity into an output signal. Specifically, MiBMI is able to read brain signals that occur when a person imagines writing a letter, and output these signals as text.

Part of the chip's success lies in a new approach to reading the language processing signals that the brain sends. While working on the chip, EPFL researchers discovered a number of very specific neural markers that are activated when a person imagines the spelling of each letter. These markers have been called "distinctive neural codes" (DNC).

DNC became a kind of shorthand for each letter, which allowed the MiBMI chipset to process exclusively the tokens themselves. The amount of data associated with each marker is about one hundred bytes, while the typical amount of neural data corresponding to the representation of a letter is measured in thousands of bytes. This reduction in the amount of processed data allowed to reduce the size of the chip and reduce power consumption. The approach will also reduce the time required to train users using the chip. MiBMI is already capable of recognizing 31 characters, which is a record for such integrated systems. The researchers plan to increase this figure to 100 characters.

The EPFL chip is designed to help people who cannot speak or move communicate with others. Researchers are now exploring possible applications of the system beyond text processing, including speech decoding and motion control. Their goal is to create a universal neurochip that can be adapted for various neurological disorders.