IEEE Computational Intelligence Magazine - February 2022 - 28

Explainable Multivariate Pattern Analysis was able to
delineate the cortical networks formed for auditory
and visual stimuli processing in sixmonth-old infants
discovering new neurocognitive patterns in the
developing brain.
concluded that drowsiness related regions are generally found
to be in parietal and occipital lobes.
For the application of ICA-FNN in DCN studies, the online
learning of the hyperparameters specific to each subject
would need to be modified, based on the task at hand, since
infants will not be able to provide feedback about their cognition
state. In addition, to address the inter- and intra-subject
variabilities, type-2 fuzzy frameworks can be utilized as
explored in these works [60], [61]. Similarly, EFCMs estimated
values of EC can be mapped generally to the specialization and
neural reuse of the DCN processes, however, not much insight
can be gained about the activation(s) of the individual cortical
regions. This is mainly because of how EFCMs seek to find the
optimal values of the EC, by trying to minimize the error
between the estimated and the actual values of the fNIRS signals
with the help of GA. Hence, not much could be inferred
about which part of the cortex is, for example, more active
from the optimized EC values.
In the next section, we review some of the AI methods as
applied to DCN studies.
B. AI in Developmental Cognitive Neuroscience
The de-facto standard for analysis of DCN studies is univariate
analysis based on simple statistical tests, where the cortical
regions most active in response to the presented stimulus is recognized,
i.e., it is an activation based analysis. There is also a
tendency of translating models used in adult research to DCN;
however, this entails making some assumptions. In contrast,
very few DCN studies have focused on decoding the multivariate
patterns in brain activity of infants in response to the presented
stimuli (such as [62] which is a correlation based
MVPA). In fact, there is an evident scarcity for undertaking AI
methods in DCN research.
In this subsection, we review the non-explainable and
explainable AI methods as applied to DCN studies for conducting
MVPA.
1) Non-Explainable AI Method
(a) MVPA with EEG using SVM
In the study by Bayet et al. [63], time resolved EEG based
MVPA is conducted using a linear SVM. Infants aged 12 to 15
months participated in the study. The aim of the study was to
investigate whether neural representations in the adult brain
are different from the developing brain for the processing of
visual stimuli (animals vs human body parts). The group-wise
classification results of the SVM based MVPA was able to suc28
IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2022
cessfully decode between infants' and adults'
brain activation patterns in response to the
presented stimuli. However, infant multivariate
representations didn't linearly separate for
animal and body images.
The study was able to establish that neural
representation for visual information processing,
of animals vs body parts differ significantly
between infants and adults. These findings were significant
by suggesting that the cortical networks undergo the processes
of localization and specialization to process the presented visual
stimulus information. However, the study could not shed
light on what cortical networks were activated for adults, and
likewise, for infants, that could explain the underlying brain
mechanism correspondingly. This is mainly because of the
non-explainable inference mechanism of SVM as discussed
previously in III-A1).
(b) MVPA with fNIRS using Correlation
In contrast to EEG's MVPA analysis (which is temporally
driven), an fNIRS based MVPA is aimed at spatial investigations
into the cortical regions' activations encoded in the
MVM. A hypothetical construction of a MVM using six
fNIRS signals reading from six different cortical regions is
depicted in Fig. 7 (a)-(b). The work by Emberson et al. [62]
decoded the brain responses in 19 six-months-old infants'
fNIRS signals in response to auditory and visual stimuli. They
decoded the signals by undertaking a MVPA driven by correlation
and reported an average classification accuracy of
66.67% for trial-level decoding.
The significance of their work lies in usage of MVPA that
improved the decoding sensitivity in comparison to their previous
work that used univariate methods [64]. A feature significance
analysis was also undertaken to determine which
features (fNIRS channels) are most significant for recognizing
the fNIRS signals in response to visual and auditory stimuli.
Their results indicated channel 1 (occipital cortex), channel 3
(occipital cortex), and channel 8 (prefrontal cortex) to be the
most critical channels for decoding between visual and auditory
stimuli.
The identification of fNIRS channels and their corresponding
anatomical locations sheds light on the localized activation
of the cortex as delineated by the IS framework. In addition,
the improved sensitivity of MVPA on account of analyzing
more than one variable (fNIRS channels' activity) rather than
univariate analysis further corroborates that cortical networks
(interaction between multiple cortical regions) are formed for
the processing of perceptual stimuli. In this sense, the correlation
based MVPA is able to implicitly imply the formation of
cortical networks. However, what exactly entails the cortical
networks is unknown because the presence and type of interaction
between the fNIRS channels is unrevealed by the correlation
based MVPA.
Motivated from the success of the correlation based MVPA
by Emberson [62] and to overcome its limitation of partial
IEEE Computational Intelligence Magazine - February 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2022
