IEEE Computational Intelligence Magazine - November 2020 - 18

Dilbag et al. [42] proposed a multi-objective differential
evolution algorithm to optimize the hyperparameters of the
regular CNN that was trained from CT images for classification of COVID-19-infected patients. A multi-objective fitness
function was designed according to both the sensitivity and
specificity of classifications of COVID-19-infected patients.
According to Dilbag et al.'s experiments, the proposed model
slightly outperformed state-of-the-art models, such as a regular
CNNs, an adaptive neuro-fuzzy inference system, and an artificial neural network. The overall improvement in terms of accuracy was 1.9789%.
According to the papers we surveyed in this section, several
insightful observations can be made:
1) Since evolutionary computation was designed for the
optimization of parameters, we can see that most works,
like [38]-[40], utilize evolutionary computation to predict virus propagation. Meanwhile, some of these works
[38], [39] adopted the concept of multi-objective genetic
algorithm to estimate the number of confirmed cases and
tackle other properties such as economic concerns. The
prediction results can also address other issues such as precaution development.
2) Although multi-objective genetic algorithm could be utilized for solving multi-objective problems, the works [37],
[38] only adopted them straightforwardly without any
modifications. Since the application scenarios of the works
of [37], [38] are different from that of multi-objective
genetic algorithms, their effectiveness is not significant in
supporting their reliability and applicability. We believe
these works could be further improved. For example, the
interaction among the multiple fitness functions could be
included to adjust for the process of optimization.

learning, which is a machine learning framework based on
mathematical analysis; 3) Vapnik-Chervonenkis theory, which is
a learning process explained by a statistical point of view; 4)
Bayesian inference, which is a statistical inference based on
Bayes' theorem; 5) algorithmic learning theory, which is a
machine learning theory explained by an algorithmic point of
view; and 6) online machine learning, which is a sort of
machine learning method for continuously updating data.
Although its primary goal is to understand learning in an
abstract manner, through the development of learning theory,
we can design various practical learning algorithms. For example, Bayesian inference is the foundation of the concept of belief
networks. Because the concept of belief networks is the foundation of the deep neural networks introduced in the previous
section, we will introduce the remaining approaches in this section, which are designed based on belief networks except deep
neural networks. Table V shows the issues that have been
addressed by existing computational learning theory methods.
Duffey and Zio [44] proposed a computational learning
theory that can learn a prediction model from the prediction
errors in the recovery time from the outbreak of the COVID-19
pandemic. This approach uses the exponential Universal Learning Curve to estimate the trend in the infection rates of the
COVID-19 pandemic. The key to the proposed approach is to
treat the infection rate as a measure of false prediction results
and time as a measure of experience/knowledge or risk exposure to allow learning. The results of Universal Learning Curve,
which was learned from China, South Korea, and other
nations, show a decreasing trajectory after a peak. The reason
might be that countermeasures are effective for controlling the
spread of the virus.
Wang et al. [45] proposed a novel noise-robust learning
framework called COPLE-Net based on the self-ensemble of
convolutional neural networks [50], [51], a sort of semi-superVI. Computational Learning Theory for
vised learning mechanism. Unlike conventional semi-supervised
Combating COVID-19
learning mechanisms that use the exponential moving average
Computational learning theory has many implementations.
of a model to adjust standard model, Wang et al. [45] developed
Based on different assumptions, various inference principles can
two designs to address the issue on noisy labels. The first design
be deduced. As a result, the deduced inference principles are
is a dynamic adjustment that can reduce the impact of the
utilized to design different computational learning theory
exponential moving average of a model while the training loss is
approaches. These approaches can usually be categorized into
decreased. The second is an adaptive learner that enables the
six types: 1) exact learning; 2) probably approximately correct
standard model to learn from the exponential moving average of a model. The
proposed COPLE-Net outperforms
TABLE IV Prediction results of Salgotra et al.'s study [41].
state-of-the-art models in terms of the
DAILY
DAILY
DAILY
DAILY
average Dice similarity (80.29%) and
CONFIRMED
DEATH
CONFIRMED
DEATH
the average 95-th percentile of HausCOUNTRY
CASES
COUNT
COUNTRY
CASES
COUNT
droff distance (18.72 mm).
USA
20,972
1358
TURKEY
1,071
17
Barstugan et al. [46] presented an
BRAZIL
28,822
1076
CANADA
717
103
early phase detection method for
RUSSIA
6,928
270
SPAIN
321
148
COVID-19 using a support vector
MEXICO
4,121
466
GERMANY
271
23
machine classifier. The classifier was
UK
3,759
204
ITALY
247
178
trained from four extensive datasets,
IRAN
1,652
57
FRANCE
191
50
which were produced by fetching patches
with sizes of 16 × 16, 32 × 32, 48 × 48,
SOUTH AFRICA
1,895
60
SINGAPORE
68
0.05

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2020

IEEE Computational Intelligence Magazine - November 2020

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2020