IEEE Systems, Man and Cybernetics Magazine - July 2020 - 13

involved in the BMI loop (i.e., the human subject and the
machine) has been expressed and termed coadaptation or
mutual learning [10], [11].
Although the need for mutual learning is widely
acknowledged in both invasive and noninvasive BMIs [12]-
[16], current research trends are heavily biased toward the
machine-learning side of BMI training. In this article, we
explore a line of reasoning promoting the reinstatement of
subject learning as an equally important pillar in BMI
training. We focus on noninvasive BMIs decoding sensorimotor rhythms (SMRs) and invasive BMIs decoding kinematics, as the evidence supporting the possibility of
learning to regulate evoked potentials remains limited [17].
Subject Learning in BMI
BMI literature is marked by a major contradiction. On the
one hand, reference subject learning, the ability of subjects
to modulate their brain activity through feedback toward
optimizing BMI performance, is ubiquitous [4]. There
seems to be a pervasive belief in the field that subject
learning is a direct, "automatic" consequence of closing
the loop between a user and a brain-actuated device. However, explicit experimental evidence is scarce, especially
for human subjects. Also, the literature that supports
human learning during BMI training and operation is
based on small cohorts [2], [8], [18]-[20].
On the other hand, the literature (overall, and specifically on BMI training) is heavily dominated by studies
oriented toward novel signal processing and machinelearning methods applied to BMIs [21]. The vast majority
of these works entail open-loop ("offline") experimentation, which forthrightly exclude any role for subject learning. Closed-loop studies using real-time feedback
("online") conducted under the umbrella of "coadaptation"
are also mainly focused on issues pertinent to the machinelearning side [22]. In a survey conducted in 2010, it was
found that less than half of the reviewed neurofeedback
and BMI publications in 50 years of relevant research,
which were longitudinal enough to qualify as learning
studies, actually reported some sort of learning effects
[23]. Even in this minority subset, in most cases, it is
unclear to what extent BMI performance improvement can
be attributed to subjects actually learning to better regulate their brain signals or other factors (e.g., machinelearning interventions).
Tackling BMI training as a purely machine-learning
problem stems from the great impact of AI on the field.
Indeed, the introduction of multivariate brain feature-processing algorithms for feature selection and decoding is
the main factor distinguishing BMIs from neurofeedback,
and critical for making it possible to control brain-actuated devices. This trend is stronger with the advent of
advanced AI algorithms like deep learning [24], [25]. Moreover, the view that ongoing developments in brain imaging
techniques [26], which promise to increase both the
amount and quality of information extracted from the
	

brain, will soon render the BMI problem a mere decoding
issue, is a popular one.
Nevertheless, despite the benefits of machine learning, a
large portion of prospective users is still unable to gain control of a BMI without adequate user training [22], [27]-[30].
The situation seems to be even more critical with regard to
people with disabilities [29], where lesions of the central and
peripheral nervous systems might affect the natural sensorimotor apparatus that machine-learning-oriented BMI
designs aim to exploit. The hope of deploying "big data" AI
approaches, however, seems limited in BMI given the need
for the rapid deployment of decoders in closed-loop interaction to keep subjects engaged and motivated. Thus far,
transfer-learning approaches also seem unable to overcome
this obstacle, largely due to the subject-specific nature of
brain patterns, which need to be employed for BMIs [31].
Finally, the promise of "zero-training" BMI and universal access to the technology for all prospective users
remain elusive. There is mounting evidence that points to
the need for shifting the focus of investigation toward subject learning and the interactions between human and
machine adaptation [12], [13], [16], [32], [33].
Pitfalls in BMI Literature on
Subject Learning
The research on subject learning in BMI is limited. In the next
sections, we identify the main caveats and pinpoint the critical improvements needed for future experimental designs.
Lack of Longitudinal Experimentation
The practice in BMI closed-loop studies seems to involve a
single or, rarely, two or three training sessions [28], [34]-
[42]. A handful of works report 5-15 sessions [29], [43]-[50],
and only a few can be characterized as truly longitudinal
studies [2], [18]-[20], [51]-[63] extending over many sessions and long periods of time. Assuming that BMI learning is, as hypothesized, a form of operant conditioning that
mainly falls under the category of implicit rather than
explicit/declarative learning, it must be expected that it
involves network- and cellular-level mechanisms akin to
associative plasticity [64]. Such biological processes cannot possibly take effect in short training periods, at least
as far as retained motor skills are concerned [65], [66]. It
can thus be said that the majority of BMI training literature has not yet fully explored the possibility to enable,
observe, and reliably evaluate subject-learning effects.
Lack of End-User Studies
Another important pitfall is the lack of end-user studies.
Although most longitudinal BMI experimentation has been
done with people in paralysis, these works comprise an
almost negligible percentage of the BMI literature. People
with motor disabilities not only represent the main target
group of BMI technology but might also be the user category for which subject learning may prove to be particularly
crucial. The pathologies of the central and even of the
Ju ly 2020

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE	

13
IEEE Systems, Man and Cybernetics Magazine - July 2020

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