IEEE Systems, Man and Cybernetics Magazine - July 2020 - 37

date, part of Inria Bordeaux-Sud-Ouest, and a member of
probably be avoided for data sets with little training data
the European Research Council BrainConquest project, aim(i.e., a few dozens). Second, filter bank RGCs (FBTSC and
ing to improve brain-computer interface (BCI) user training.
FBFgMFM) should also be recommended to obtain good
His research is focused on measuring learning-related mental
classification performances, notably with subject-specific
states through electroencephalocalibration, for both workload and
graphic and physiological signals,
emotion classification, whatever
investigating machine learning
the amount of training data. Howalgorithms for decoding such sigever, such methods do not seem to
CNN and the
nals, BCIs, human-computer interbe suitable for subject-independent
proposed filter bank
action, machine learning, and cogclassification with little training
RGCs are valuable
nitive science.
data or for emotion classification.
Andrzej Cichocki (A.Cichocki@
They do seem to be suitable for the
machine-learning
skoltech.ru) earned his M.Sc. (with
subject-independent classification
tools for scientists
honors), Ph.D., and Dr.Sc. (Habilitaof workload with a large amount of
tion) degrees, all in electrical engitraining data, however.
who are decoding
neering, from Warsaw University of
Our results also confirmed that
cognitive and
Technology, Poland. He was with
passive BCIs with subject-indepenaffective states from
University Erlangen-Nuerenberg,
dent calibration are possible but
Germany, as an Alexander-vonare very challenging and feature
EEG signals.
Humboldt research fellow and guest
much lower accuracies. Similarly,
professor. From 1995 to 2018, he
affective state classification in
was a team leader and the head of
EEG is possible but is much more
the Advanced Brain Signal Processing Laboratory at RIKEN
challenging than workload estimation. However, those
Brain Science Institute in Tokyo. Under his guidance, the
suggestions imply computational costs that will differ
new Tensor Networks and Deep Learning for Applications in
from one algorithm to another. Using the FB RGCs or CNN
Biomedical Data Mining Laboratory was established at
will require a long calibration time when the testing phase
Skoltech, Moscow. His publications currently have over
may also be time consuming and must be considered
42,000 citations, with an h-index of 93 according to Google
before moving toward online uses. See the supplementary
Scholar. He is a Fellow of the IEEE.
material for more information.
Fabien Lotte (fabien.lotte@inria.fr) earned his M.Sc.,
For the emotion data set, we labeled trials as in the
M.Eng., and Ph.D. degrees in computer sciences, all from
original article to allow comparisons, i.e., with a global
the National Institute of Applied Sciences Rennes, France,
subject-independent partition between low/high valence/
in 2005 (M.Sc. and M.Eng.) and 2008 (Ph.D.), respectively,
arousal trials, based on the Self-Assessment Manikin ratand his habilitation degree in computer science from the
ings. Note that better methods to partition low/high trials
University of Bordeaux in 2016. Since October 2019, he has
in a per-subject basis can also be used in the future [42],
been a research director at Inria Bordeaux-Sud-Ouest,
to limit the class imbalance, and other deep learning
France, on the Potioc team. His Ph.D. thesis received both
architectures, notably recurrent neural networks [34],
the Ph.D. Thesis Award from the French Association for
may prove to be promising for EEG classification and pasPattern Recognition in 2009 and the Ph.D. Thesis Award
sive BCIs as well. It would also be interesting to study
accessit (second prize) in 2009 from the French Association
whether CNN and RGCs can be used to robustly estimate
for Information Sciences and Technologies. He is part of the
other cognitive states such as fatigue, curiosity, or engageeditorial boards of Brain-Computer Interfaces (since 2016)
ment, and how well the proposed RGCs perform on motor
and Journal of Neural Engineering (since 2016). In 2016,
imagery data for active BCIs. Altogether, our results sughe was the recipient of a European Research Council Startgested that CNN and the proposed filter bank RGCs are
ing Grant (project BrainConquest) to develop his research
valuable machine-learning tools for scientists who are
on brain-computer interfaces.
decoding cognitive and affective states from EEG signals.
Acknowledgments
This work was supported by the European Research Council under grant ERC-2016-STG-714567 and by the Japanese
Society for the Promotion of Science.

References
[1] M. Clerc, L. Bougrain, and F. Lotte, Brain-Computer Interfaces 1: Methods and
Perspectives. Hoboken, NJ: Wiley, 2016.
[2] J. R. Millán et al., "Combining brain-computer interfaces and assistive technologies: State-of-the-art and challenges," Front. Neurosci., vol. 4, p. 161, Sept. 2010.

About the Authors
Aurelien Appriou (aurelien.appriou@inria.fr) earned his
M.Sc. degree (with honors) in cognitive science from the
University of Bordeaux, France, in 2015. He is a Ph.D. candi

doi: 10.3389/fnins.2010.00161.
[3] T. O. Zander and C. Kothe, "Towards passive brain-computer interfaces: Applying brain-computer interface technology to human-machine systems in general,"
J. -Neural. Eng., vol. 8, no. 2, p. 025005, 2011. doi: 10.1088/1741-2560/8/2/025005.

Ju ly 2020

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE

IEEE Systems, Man and Cybernetics Magazine - July 2020

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - July 2020