A system has been developed to create smarter brain-computer interfaces

A system has been developed to create smarter brain-computer interfaces
A system has been developed to create smarter brain-computer interfaces

Experts from Skoltech, INRIA and RIKEN Advanced Intelligence Project investigated the capabilities of several modern machine learning algorithms to solve the determination of mental load and affective states of a person. The developed software can be used to create smarter brain-computer interfaces (BCIs), which can be used in medicine and other fields.


The research results are published in the journal IEEE Systems, Man, and Cybernetics. A BCI provides a link between the human brain and a computer, allowing a person to control various devices, such as a robot arm or a wheelchair, based on a signal from the brain (active BCI).

BCI also allows you to track and classify psycho-emotional states of a person (passive BCI). The signals from the brain to the BCI are usually measured using electroencephalography (EEG), a common non-invasive method for measuring the electrical activity of the brain.

The "raw" data obtained as a result of EEG in the form of continuous signals must undergo sufficiently thorough processing before they can provide an accurate determination of the mental load and affective states of a person, which is a prerequisite for the correct operation of a passive BCI.

The experimental data available to date indicate that the accuracy of these measurements is insufficient even for solving such simple problems as determining the difference between high and low mental workload, not to mention their use in practical applications.

“This low measurement accuracy is due to the extremely complex structure of the human brain. Imagine that our brain is a huge orchestra, which involves thousands of instruments, and we need to use a limited number of microphones and sensors to highlight the characteristic sound of each individual instrument,”said one of the authors of the article, professor at the Skoltech Center for Scientific and Engineering Computing. technologies for tasks with large amounts of data (CDISE) Andrzej Chychocki.

It follows from this that more reliable and accurate algorithms are required to solve the problems of classifying EEG data and recognizing various patterns of the brain. Professor Chikhotskiy and his colleagues examined two groups of machine learning algorithms, Riemannian geometry classifiers (RGC) and convolutional neural networks (CNN), which performed well in active BCIs.

The researchers decided to find out whether these algorithms will cope not only with the so-called imaginary motor tasks, in which the subject imagines certain movements of the limbs, without actually performing them, but also with the tasks of assessing mental load and affective states.

Scientists held a kind of "competition" for seven algorithms, two of which scientists developed independently by optimizing well-proven Riemannian algorithms. In one of the two experiments, a typical BCI scheme was used, in which the algorithms were first trained on data about a particular subject, and then tested on that subject.

The second experiment was carried out without reference to a specific subject, and this scheme is much more complicated, since the activity of the brain can be very different for different people. The experiments used real EEG data from earlier experiments by one of the authors of the article Fabien Lotte and his colleagues, as well as the DEAP database, which collected data on the analysis of human emotional states.

Scientists found that the deep neural network outperformed all its "competitors" in solving the problem of assessing mental load, but at the same time it did a poor job of classifying emotional states, but two algorithms with Riemannian optimization performed well in solving both problems.

In the article, the authors conclude that it is much more difficult to use a passive BCI for classifying affective states than for assessing mental load, and the calibration of the algorithm without reference to a particular subject still gives significantly lower accuracy.

“In the next stages of the research, we plan to use more sophisticated methods based on artificial intelligence (AI) and, first of all, methods of deep learning, which can be used to detect the smallest changes in the signals and patterns of the brain.

Deep neural networks can be trained on large datasets containing information about a large number of subjects, various test scenarios, and test conditions. Artificial intelligence, the creation of which has become a real revolution, can be very useful for BCI and solving problems of recognizing human emotions, "Chikhotsky said.

Popular by topic