IEEE Computational Intelligence Magazine - February 2022 - 22

however, to what extent the connectivity analysis effectively
contributes to the understanding of brain processes is dependent
on the choice of the AI technique (used for the connectivity
analysis).
2) Representation Learning
Many recent works in neuroscience are increasingly using deep
leaning paradigms to investigate the underlying brain activity in
response to a presented task [43]. Amongst Deep Neural Networks
(DNNs), convolutional neural networks (CNNs) have
gained particular interest because of their remarkable performance
in unsupervised automatic feature extraction and classification
of objects in challenging image classification problems
[43]. Owing to the capability of CNNs to compose higher
level features using lower level features, CNNs can learn representations
of input data automatically overcoming the long
standing challenge to handcraft a feature set in conventional AI
methods [43]. In CNNs, a small matrix of numbers (called a
filter) is passed over (convoluted with) the raw data, to extract
features from the raw data, such as edges in images, also called a
feature map. The convolution layer is followed by a pooling
layer which downsamples the input to reduce both the spatial
size of the input data and the number of hyperparameters in
the network. A typical CNN architecture consists of the following
stages:
i) Feature Learning Blocks
* Convolution (C) + Rectified Linear Unit (ReLU).
* Pooling (P).
ii) Classification/Regression Blocks
* Fully Connected Layers.
* Softmax, Logistic regression layer, regression loss (Root
Mean Square Error (RMSE) etc.)
The performance of CNNs is critically dependent on the
optimization of hyperparameters, and owing to the large number
of hyperparameters that need optimization, most DNNs,
including CNNs, require large datasets to converge. The hyperparameters
of a CNN include the size of the filter(s), stride,
number of hidden layers, and the learning rate.
3) Multivariate Pattern Analysis
In most multivariate analysis, the feature set is crafted by hand,
i.e., the statistic characteristic (such as the mean or amplitude)
of a neuroimaging signal which would best capture the neural
underpinnings, corresponding to the task at hand, is chosen
manually. The two dimensional matrix formed by collating
together the features from N channels (for fNIRS) or electrodes
(for EEG) and J number of data trials is then given as
input to an AI method, and is hereby referred to as a multivariate
matrix (MVM).
Although it requires considerable subject-matter expertise
to curate a feature set for MVM that best represents the underlying
neural activity, the classification results based on the analysis
of MVM would reflect on the representational dynamics of
the underlying cortical networks (as read from fNIRS channels
or EEG electrodes). In this regard the classification results
22 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2022
obtained from the analysis of MVM can be at least partially
attributed to the cortical networks activation as represented by
the statistical feature used for constructing the MVM.
The MVM can be readily analyzed using any state-of-theart
AI methods. Most AI methods such as Support Vector
Machine (SVM) and Random Forest (RF) usually give very
robust classification results with MVM. This analysis approach
is termed multivariate pattern analysis (MVPA) [44] and was
first used for neuroimaging analysis on adult multi-voxel
fMRI data [45].
In the following subsections, we review the explainable and
non-explainable AI methods used on the aforementioned analysis
paradigms on both non-developmental (adults) and developmental
(infants) population.
A. AI in Cognitive Neuroscience for Adult Brains
In Cognitive Neuroscience, AI methods are frequently used
with adult populations (mature brains). Some approaches can
provide no explanation or simply partial information, and others
can derive some explainable structure.
1) Non-Explainable AI Methods
A review of the non-explainable AI methods for investigating
cognitive processes in adults' cognitive neuroscience studies is
presented next.
a) FC with EEG using SVM
The EEG studies by Moezzi et al. [46] and Klados et al. [47]
used SVM with radial basis function (RBF) as the underlying
kernel to investigate EC. In particular, the work by Moezzi
et al. is of interest with respect to DCN research as it investigated
the difference in FC to recognize young (mean age 24
years) from old adult brains (mean age 71 years). The FC was
studied in the standard frequency bands of delta (1-4 Hz), theta
(4-8 Hz), alpha (8-13 Hz), beta (13-30 Hz) and gamma (30-
45 Hz). The aim was to study oscillations in standards frequency
bands to uncover coordinated activity in large-scale brain
networks which facilitate information flow between spatially
distributed brain regions. The calculation of FC matrices was
done using imaginary coherence in an attempt to account for
the poor spatial localization of the EEG signals.
Cross-validation was performed to optimize the hyperparameters
(C: regularization factor, and gamma kernel coefficient)
of SVM, improve accuracy, and identify the most
significant features. To map the FC to brain regions, a grouping
approach was used to spatially localize the observed connectivity
patterns. In addition, consensus features were obtained using
Euclidian distance between electrode pairs to investigate FC
patterns based on age. They concluded that consensus features
belonging to delta, theta, alpha and gamma frequency bands
had positive weights showing significantly higher FC in younger
adults than in older adults. Features of the beta band had
negative weights showing significantly higher functional connectivity
in older adults than younger adults. However, as is also
acknowledged in the original study [46], the limitation to map
IEEE Computational Intelligence Magazine - February 2022

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