IEEE Systems, Man and Cybernetics Magazine - July 2020 - 14

peripheral nervous system disrupt the normal action-perception loop of human-environment interaction and are
known to often negatively influence the usual cortical
activity patterns elicited by able-bodied individuals [29],
[34], [39]. In other words, in this critical user group, ma--
chine learning alone seems unlikely to lead to universal
access to BMI because the "decodable" brain activity patterns that an AI model could learn to interpret may no longer be there at all; instead, subject learning and its
accompanying plasticity might be needed to create "de
novo" neuronal circuits that the BMI algorithm can later
successfully translate into actions of an end-effector [12].
Shortcomings in the Quantification
of Subject Learning
Subject learning in BMI has not been properly assessed
thus far in the literature. A minimal quantitative piece of
information that should be provided is the evolution of
BMI proficiency metrics over time as, based on psychological studies and general learning literature, the demonstration of learning curves is a "sine qua non" prerequisite to
establish the existence and typology of learning effects.
Yet, many works discuss subject learning without reporting such learning curves [19], [27], [29], [36], [54].
Extrapolating From Other Fields
A reason for claiming subject learning in BMI without hard
evidence is its similarity to human neurofeedback experimentation. Although such a resemblance exists [67], there
are good reasons why these arguments fall short. In particular, the extrapolation from neurofeedback to BMI learning is a fairly farfetched one. As previously noted, the main
difference is that neurofeedback requires users to learn to
regulate univariate brain activity; for example, the amplitude or the power in some frequency band of a single electroencephalogram channel.
Conversely, a modern BMI is driven by the output of a
classification or regression model that combines several
spatiospectrotemporal features of brain activity, even if
the actual controlled effector is unidimensional (i.e., a
visual feedback bar moving left/right or up/down). Hence,
the learning burden in a BMI seems to be much heavier
[66]. It is currently unclear how humans can cope with this
situation, whether they may be overwhelmed and give up,
how they can subconsciously focus and gain control over a
learnable manifold [68], or whether they prove able to gradually exploit the whole available multivariate feature
space. In addition, even multimodal neurofeedback studies
are simplistic when compared to the complex end-effectors and surrounding environments (and the interactions
therein) a user must learn to master a BMI, especially in
real-world conditions [20]. From this perspective, BMI
learning seems to be much closer to natural motor and
skill learning [6], [8] than to neurofeedback.
Another overstated extrapolation derives from the wellsubstantiated subject-learning effects shown in invasive
14	

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE Ju ly 2020

BMI studies with animal models [3], [5]-[7], [13]. Despite the
great value of these works in understanding BMI learning
and control, especially taking into account the seminal and
holistic manner in which the issue of subject learning has
been quantified for regular [6] and coadaptive [13] BMI systems, the translational value of subject learning in BMI
must be experimentally verified: the need to transition
from animal model studies to human clinical trials to prove
the effectiveness of a therapy or medical device is a concept well established in the medical world, yet somewhat
overlooked in the engineering-oriented BMI field. Even the
recent longitudinal studies with end users that sparked a
renewed interest in BMI and can be thought to be the natural successors of the aforementioned previous efforts with
nonhuman primates or rodents, still largely neglect the
issue of subject learning [53], [56], [59], [61], [62], save a few
notable exceptions [57]. Moreover, the great differences in
hardware, brain signals, or features and control paradigms
employed in this line of research make it unclear whether
subject learning occurs similarly in other types of BMIs.
Flaws in the Quantification of Subject Learning
Although there exist studies providing quantitative evidence for the emergence of human-learning effects during
BMI training and operation, their methodologies are open
to criticism. The most typical weakness is reporting only
the learning curves of either BMI application performances [50], [69] or the output from the BMI decoder [29], [35],
[36], [45], [47], [52], [56]. The problem is that such "high-level" metrics cannot disentangle the individual contributions
of the subject and the BMI decoder from the overall performance, effectively ignoring the presence and individual
roles of the two learning agents in the BMI loop.
We suggest that BMI control emerges in a vertical/hierarchical, rather than horizontal, fashion (see Figure 1). The
subject is at the base of the pyramid (first level) and must
necessarily be able to generate distinct patterns of brain
activity in some physiological brain feature space. Subjects will subsequently learn to improve these brain patterns with practice and the support of the other two higher
levels. In the second level, machine-learning techniques
identify the brain activity feature space where subjects'
intentions are optimally distinguishable (feature extraction and selection) and build an optimal multivariate
decoder (classification/regression). In the third level, the
output of the decoder is mapped to actions of the braincontrolled device. This mapping can benefit from shared
control approaches to increase the reliability of the device
and reduce the subject's workload, thus facilitating subject
learning [58], [70], [71].
As becomes apparent in this abstraction, reporting
only the BMI performance (second level) or the application performance (third level) cannot isolate and assess
subject learning (first level). For instance, a random classification outcome could be attributed either to the subject not being able to produce separable patterns or to the



IEEE Systems, Man and Cybernetics Magazine - July 2020

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

Contents
IEEE Systems, Man and Cybernetics Magazine - July 2020 - Cover1
IEEE Systems, Man and Cybernetics Magazine - July 2020 - Cover2
IEEE Systems, Man and Cybernetics Magazine - July 2020 - Contents
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 2
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 3
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 4
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 5
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 6
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 7
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 8
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 9
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 10
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 11
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 12
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 13
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 14
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 15
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 16
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 17
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 18
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 19
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 20
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 21
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 22
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 23
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 24
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 25
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 26
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 27
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 28
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 29
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 30
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 31
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 32
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 33
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 34
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 35
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 36
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 37
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 38
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 39
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 40
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 41
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 42
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 43
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 44
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 45
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 46
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 47
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 48
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 49
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 50
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 51
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 52
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 53
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 54
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 55
IEEE Systems, Man and Cybernetics Magazine - July 2020 - 56
IEEE Systems, Man and Cybernetics Magazine - July 2020 - Cover3
IEEE Systems, Man and Cybernetics Magazine - July 2020 - Cover4
https://www.nxtbook.com/nxtbooks/ieee/smc_202110
https://www.nxtbook.com/nxtbooks/ieee/smc_202107
https://www.nxtbook.com/nxtbooks/ieee/smc_202104
https://www.nxtbook.com/nxtbooks/ieee/smc_202101
https://www.nxtbook.com/nxtbooks/ieee/smc_202010
https://www.nxtbook.com/nxtbooks/ieee/smc_202007
https://www.nxtbook.com/nxtbooks/ieee/smc_202004
https://www.nxtbook.com/nxtbooks/ieee/smc_202001
https://www.nxtbook.com/nxtbooks/ieee/smc_201910
https://www.nxtbook.com/nxtbooks/ieee/smc_201907
https://www.nxtbook.com/nxtbooks/ieee/smc_201904
https://www.nxtbook.com/nxtbooks/ieee/smc_201901
https://www.nxtbook.com/nxtbooks/ieee/smc_201810
https://www.nxtbook.com/nxtbooks/ieee/smc_201807
https://www.nxtbook.com/nxtbooks/ieee/smc_201804
https://www.nxtbook.com/nxtbooks/ieee/smc_201801
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_1017
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0717
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0417
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0117
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_1016
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0716
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0416
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0116
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_1015
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0715
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0415
https://www.nxtbook.com/nxtbooks/ieee/systems_man_cybernetics_0115
https://www.nxtbookmedia.com