IEEE Systems, Man and Cybernetics Magazine - January 2022 - 23

WR IR WSRS
T
T uu d
R##
Dd
,
SS I() .
12
+=
== !!RDD
RR
,
the preceding CNN models are adapted and fine-tuned for
regression. We implement three models: the shallow CNN,
deep CNN, and EEGNet.
Optimization is conducted on an orthogonal manifold; i.e.,
RR .IT
=
nal matrix R:
RR IR
*
=-aa+
R
s
d
argmin() () (),JJ1
where dD :1 for the subspace method and d 1= for the
sequential optimization or deflation method.
:
CNNs for Regression
In a nutshell, a CNN is a multilayer feedforward NN
designed to learn spatial dependencies by employing
fundamental subsystems as convolution layers, pooling
layers, and fully connected layers. Convolution and pooling
layers extract features, while fully connected layers
transform the computed features to an output, which is
useful for classification and regression. Recently, several
studies developed novel models for CNNs adapted to
motor imagery classification and P300 classification in
EEGs, including the shallow CNN, deep CNN [27], and
EEGNet [31]. A deep CNN is composed of four convolution
maximum pooling blocks, with a specified primary
block outlined to work with an EEG input, accompanied
by three typical convolution maximum pooling clusters
and a dense softmax classification output. The idea of
split convolutions was used in the first convolution cluster.
It consisted of a temporal convolution followed by
spatial filtering. This was followed by maximum pooling
and subsequent convolutional layers. Linear units were
used in the temporal convolution, and exponential linear
units were employed in spatial convolution. A shallow
CNN has three layers and tunable parameters. It was
recently tested and validated for classification problems
[27]. The preliminary layer performs convolution across
the time direction, while the subsequent layer accomplishes
convolution across the spatial dimension, i.e.,
across EEG channels.
Convolution across the time dimension focuses on optimizing
bandpass filters, and spatial convolution seeks to
optimize spatial filters. The obtained signal amplitude is
squared and averaged in a pooling fashion to derive the
band power. Then, the last one is a fully connected linear
classification layer. Although the network processes the
signal in a manner analogous to the FBCSP, there is a difference
in terms of the convolutional network conjointly
optimizing spatial and temporal filters. In summary, this
CNN processes EEG data in a manner similar to the
FBCSP and linear discriminant analysis. In contrast to the
FBCSP, all these filters are simultaneously optimized, producing
better performance using motor EEG signals. A
shallow CNN uses minimally preprocessed EEG signals as
input, so we filtered the signals at 4-40 Hz. In this article,
Implementation and Results
The objective functions now depend on orthogoPAT
and EEG Preprocessing
The PAT experiment is shown in Figure 3. The goal is to
study the correlation between fatigue and driving performance,
based on the proposition that poor vigilance leads
to significant delays before drivers notice events. The
experiment details, EEG trial preprocessing, and RT processing
remain the same as reported in [5, Secs. 3(a), 3(b)
(1), and 3(b)(2)]. The raw data are available for download
from https://doi.org/10.6084/m9.figshare.6427334.v5 [3].
They are targeted to assess RTs by utilizing a 5-s EEG window
shortly ahead of drowsiness/alert states.
Methods for Performance Comparison
In practice, three approaches-an EEGNet [31], a shallow
CNN [23], and a deep CNN [27]-applied to EEG trials are
generally used for predicting RTs. The fuzzy CSP [18] and
FTDCSSP [5] feature extraction methods with LASSObased
RT prediction are additional methods used as baselines.
We compare the performance of the proposed
methods (F-DivCSP-WS- and F-DivIT-JAD-based feature
extraction with LASSO-based RT prediction), with the preceding
approaches used as baselines.
Hyperparameter Tuning and
Performance Comparison
Figure 3 and 4 along with Figures 5 and 6 in the supplementary
material depict the average performance of the
methods on the RT data set. The analysis is done with a
leave-one-session-out validation approach for each of
the subjects. Subjects
SSSSSS S
04 06 11 23 48 52 54 ,,,,,, ,
and S55 are left out of the analysis, as their data were
limited to a single session. The EEGNet, deep CNN, and
shallow CNN architectures are trained via the Keras
deep learning platform with a TensorFlow backend, and
the server is an Intel Xeon CPU with a 2.2-GHz processor,
a Tesla T4 15079MiB GPU, and 13 GB of randomaccess
memory. Five hundred epochs were used to train
the networks.
Hyperparameter Tuning in Each Method
Shallow CNN
The shallow CNN was the slowest, with an epoch taking
5-6 s. This is attributed to the custom activation functions
that were implemented from scratch and not optimized for
the model. To reduce the time complexity, we changed the
filters in the first convolution layer to 20, from the original
40. The learning rate obtained from the line search was
0.0001, and the average model showed better results. The
network had approximately 26,500 trainable parameters.
January 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 23
https://www.doi.org/10.6084/m9.figshare.6427334.v5
IEEE Systems, Man and Cybernetics Magazine - January 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - January 2022
