IEEE Systems, Man and Cybernetics Magazine - July 2020 - 16

participants. Therefore, there is considerable room for
improvement, which we strongly believe will also help reveal
the true potential of subject learning for enabling people to
control BMI-based applications. In the following sections, we
discuss the most important open issues emerging from the
study of BMI learning and coadaptation literature with reference to lessons learned from our own previous work.
Nature of Subject Learning
Subject learning in BMI has been mainly discussed in the
contexts of operant conditioning (instrumental learning),
motor learning (especially regarding systems involving
motor-related mental tasks like invasive BMI, decoding kinematics, and noninvasive SMR-based BMI), and general skill
learning. Given the very broad definitions of all these frameworks, the abundance of different methodologies each one of
them encompasses and the variety of different BMI paradigms, these categorizations should rather be viewed as
complementary and overlapping, not as mutually exclusive.
For instance, the role of feedback contingent to entrained
brain activity is regarded as crucial under all these schemes.
One of the most important questions regarding the
nature of subject learning in BMI pertains to whether it is
mostly an implicit/procedural or explicit/declarative process. Implicit learning suggests that subjects may gain control over BMIs gradually, in a largely subconscious manner
through feedback observation, as is shown to occur in neurofeedback studies. This mode of learning is more compatible with the instrumental and skill-learning theories [67].
On the other hand, explicit learning relies on declarative
knowledge passed on to the subject verbally or schematically, i.e., through instructions, examples, and illustrations.
Delineating the implicit and explicit aspects of learning
in BMI is crucial for a number of reasons. First, it is a critical factor in training protocol design. For example, the
explicit learning approach calls for more "explanatory"
protocols that better take into account the educational theories of learning or, the instructional and motivational
designs [32], whereas implicit learning could more likely
be boosted by interventions such as the establishment of
more natural feedback provision strategies [74]-[76]. Second, the expected timescales of learning (and, as a result,
the required training times) should depend heavily on the
degree of its explicit or implicit nature. Specifically, declarative learning paradigms are more likely to induce abrupt
changes in performance and in the elicited task-dependent
brain patterns, which are coupled to the onset of adopting
a new successful (i.e., separable) mental task or strategy.
These changes should probably reflect the preexisting
engrams of the newly employed mental tasks rather than
any functional plasticity and cortical reorganization.
On the contrary, implicit learning is thought to rise
from the plastic effects encoding a newly acquired skill
and thus should more likely manifest with the smooth and
gradual emergence of brain patterns and performance
improvements that follow some associated functional and/
16	

IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE Ju ly 2020

or structural plasticity. The only principled investigation
conducted so far on the issue of the (procedural or declarative) nature of learning in SMR-based BMI has argued in
favor of the implicit learning model [77]. Also, the fact that
proficient BMI users often report reaching a state of "automaticity" after long-term use, where they no longer explicitly employ the instructed "surrogate" mental task (e.g.,
motor imagery of some limb) but instead directly command the control of the BMI actuator [10], [19], [20], [44],
[54], [55], also supports an implicit learning process.
Time Scales of Subject Learning
Although most studies imposing from five to 15 training sessions [29], [43]-[50] (including our own work with end users
[29], [46], [49]) have mostly failed to show clear indications
of subject learning, our recent study involving two users
with a spinal cord injury has shown that people with severe
disabilities can learn to operate an SMR BCI in real-world
conditions with a slightly higher number of sessions (15-20)
[20]. Importantly, they started training with a poor ability to
spontaneously modulate SMRs and outperformed other participants with similar disabilities in an international competition. Even though our subjects relied on a mutual-learning
approach that explicitly elicited subject learning, all the
other participants seemed to follow a conventional
machine-learning approach, including multiple feature reselection and classifier recalibration rounds. It, therefore,
appears that, BMIs adequately supported by machine learning that also foster subject learning may deliver better longterm outcomes. The few longitudinal studies in which BMI
subjects received only sparse or parsimonious machinelearning interventions have led to comparable conclusions
[55], [57]. Also, these are approximately the derived timescales of learning in early animal studies [3], [5].
Assuming that this amount of training sessions is a minimum to induce subject learning, the question is raised
whether claiming BMI learning within a single or a couple
of sessions merits any scientific grounding. It is well established that the acquisition of motor and general skills
evolves with an initial "fast-learning" (even intrasession)
phase followed by a "slow-learning" (multisession) one [67],
[78]-[80]. Both stages have been linked with functional synaptic plasticity in animal models and humans [79]. Hence,
some form of BMI learning taking place in short time scales
cannot be excluded; however, neural circuit changes in the
fast-learning phase have been thought to reflect short-term
plasticity, i.e., effects tend to return to baseline in a matter
of minutes or hours. Furthermore, motor skill consolidation and, eventually, retention, has been shown to require
long-term training and leads to larger cortical reorganization or even structural plasticity.
Redefining Coadaptation: When and What
Machines and Humans Learn
As with the rest of the BMI literature, the studies introducing coadaptation in BMI have also almost exclusively
IEEE Systems, Man and Cybernetics Magazine - July 2020

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