Researchers at Skoltech, INIAA, and the Riken Advanced Intelligence Project have considered several state-of-the-art machine learning algorithms for the challenging tasks of determining the mental workload and emotional states of the human brain.
Their software can help design better brain-computer interfaces for applications in medicine and beyond. The paper was published in the journal IEEE Systems, MAN and Cybernetics.
A brain-computer interface, or BCI, is a link between a human brain and a machine that can allow users to control various devices such as robotic weapons or wheelchairs only by brain activity (these are called active BCIs Is) or can monitor the user’s mental state or emotions and classify them (these are passive BCIs). Brain signals in BCI are usually measured by electroencephalography, a non-common method of recording the brain’s electrical activity.
But there is a long way from raw continuous EEG signals to digitally processed signals or patterns that have the ability to accurately identify a user’s mental workload or affective states, something that passive BCI needs to make functional. Existing experiments have shown that the accuracy of these measurements, even for simple tasks discriminating from high workloads, is insufficient for reliable practical applications.
“The low accuracy is due to the extremely high complexity of the human brain.
The brain is like a giant orchestra with thousands of musical instruments, from which we want to extract the specific sounds of each individual instrument using a limited number of microphones or other sensors, ”Andrews Schiecki, Skoltech Center in Computational and Data-intensive Science Professor and Engineering (CDISE) and a base of paper, notes.
Thus, more robust and accurate EEG classification and validation of various brain pattern algorithms is badly needed. Cichoki and his colleagues looked at two groups of machine learning algorithms, the Rimanian geometry-based classifier (RGC) and the Conventional Neural Network (CNN), which are performing quite well on the active side of the BCI.
Researchers wondered whether these algorithms could work not only with so-called motor imaginary functions, where a subject imagines certain movements of the limbs without any real movement, but for workload and affective approximations.
He ran a competition of sorts among seven algorithms, two of which scientists designed themselves by improving well-performing Rimanian methods.
The algorithm was tested in two studies, one with a specific arrangement for the BCI where the algorithm was trained on data from a specific subject and subsequently tested on the same subject, and one that was subject-independent – A much more challenging setup, waves can vary greatly since our brain.
Real EEG data were taken from earlier experiments conducted by Fabien Lotte, a co-author of the paper, and his colleagues, as well as from DEAP, an existing database for sentiment analysis.
For example, scientists found that an artificial deep neural network significantly outperformed all of its competitors in the workload estimation task, but performed poorly in emotion classification. And both modified Reimanian algorithms did quite well in both tasks.
Overall, as the paper concludes, using passive BCI for lubricant state classification is much harder than estimating workload, and leads to subject-independent calibration, at least for now, to much lower accuracy .
“In the next steps, we plan to use more sophisticated artificial intelligence (AI) methods, especially intensive learning, which allow us to detect very small changes in brain signals or brain patterns.
Deep neural networks can be trained based on a large set of data from multiple disciplines in different scenarios and under different circumstances. AI is a real revolution and is also potentially useful for BCI and recognition of human emotions, ”said Sichoki.
Estimating cognitive or affective states from brain signals is an important but challenging step in creating passive brain – computer interface (BCI) applications. Until now, it has been possible to assess mental workload or emotions from electroencephalogram (EEG) signals with only modest classification accuracy, thus giving rise to unreliable neurodeptive applications.
However, recent machine-learning algorithms, notably the Rimanian geometry-based classifier (RGC) and the Confusional Neural Network (CNN), have shown promise for other BCI systems, such as, motor imagined BCI. However, they have not been formally studied and compared to classifications of cognitive or affective states.
