WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,

DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

Artificial Intelligence: Learning and Limitations

ALISSON PAULO DE OLIVEIRA1 HUGO FERREIRA TADEU BRAGA2

Graduate Program in Metallurgical, Materials, and Mining Engineering

Universidade Federal de Minas Gerais1
Av. Antônio Carlos, 6627 - Campus da UFMG – Pampulha, Escola de Engenharia, BRAZIL.

Innovation Center2
Fundação Dom Cabral
Avenida Princesa Diana, 760, Alphaville, Lagoa dos Ingleses, BRAZIL.
Received: June 5, 2020. Revised: July 17, 2020. Accepted: July 21, 2020. Published: July 22, 2020.

Abstract: Artificial Intelligence, IA, is a new technology with enormous potential to change the world forever as we know
it. It finds applications in many fields of human activity, including services, industry, education, social networks,
transportation, among others. However, there is little discussion about the accuracy and reliability of such technology,
which has been used in situations where human life depends on its decision-making process, which is the result of its
training, one of the stages of development. It is known that the learning process of an Artificial Intelligence, which can use
the Artificial Neural Networks technology, presents an error of the predicted value in relation to the real value, which can
compromise its application, being more critical in situations where the user's security is a major issue. In this article, we
discuss the main technologies used in AI, their development history, considerations about Artificial Neural Networks and
the failures arising from the training and hardware processes used. Three types of errors are discussed: The Adversarial
Examples, the Soft Errors and the Errors due the lack of Appropriate Training. A case study associated with the third type
of error is discussed and actions based on Design of Experiments are proposed. The objective is to change the way the AI
models are trained, to add some rare conditions, and to improve their ability to forecast with greater accuracy in any
situation.

Keywords: Artificial Intelligence; Artificial Neural Networks; Deep Learning; Machine Learning; Adversarial Examples;
Soft Errors.

1. Introduction where an AI system suggested a wrong treatment for a

cancer patient what could worsen his condition [3], an AI
The Artificial Intelligence, AI, is a technology related to chatbot assuming unwanted "personalities" and assuming
the ability of a digital computer to perform tasks the worst tendencies from the Internet [4], among others.
commonly associated with intelligent beings. The term is Some types of errors (Adversarial examples, soft errors
often applied to systems development projects supported and errors due the lack of appropriate training) are
with human-characteristic technological processes, such discussed. This paper focus, as major contribution, on a
as reasoning skills, meaning discovery, or learning from specific type of error originated from the lack of
past experiences. Thus, such technology can replicate appropriate training. This error is discussed with a real
human behavior in matters related to the interpretation of problem faced during the development of a prediction
situations (duly converted into data) and subsequent model of steel properties. It is known [5, 6] that the error
decision based on the generation of information. in the training of artificial intelligence algorithms is
Moreover, this technology proves to be able to predict reduced with large availability of data regarding a
future situations from past experiences. However, the situation where an AI with minimal error is desired. So,
accuracy (or error obtained from an actual value) is far necessarily, regions with little existing data lead to poor
from being desired for critical applications where the user AI learning with considerable response errors. This
safety is a key factor [1]. article seeks to explore the learning mechanism on which
Artificial Intelligence is based and its main limitations. A
Today, there is a boom in the development of a variety of relationship between the error magnitude and the amount
artificial intelligence applications, from image of data available to training is also discussed and possible
recognition to even self-driving cars, to the suggestion of mitigation actions are suggested with the objective to
medical treatments. Unfortunately, however, there are reduce this error by designing an appropriate experiment
cases of failure of an autonomous driving system where to generate data with low frequency.
apparently it was not unable to identify when to stop the
and this lead to an accident with loss of life [2], a case

E-ISSN: 2224-3410 80 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

2. Overview of Artificial Intelligence

 Manufacturing robots;
Technologies  Autonomous driving system;
 Smart assistants;
2.1. History of Artificial Intelligence  Social media monitoring;
 Natural Language Processing (NLP) tools;
Recently, fast progress has been made in the fields related  Proactive healthcare management;
to Neuroscience and Artificial Intelligence (AI). At the  Conversational marketing bots;
beginning of the computer age, work on artificial  Prediction models;
intelligence was interrelated with neuroscience and  Image reconnaissance tools.
psychology, and many of the pioneer authors traveled
both fields, with collaborations between these disciplines There are many other applications of AI technology.
proving to be very productive [7]. More recently,
however, this interaction has become much less common,
as both subjects have grown a lot and disciplinary
2.2. Artificial Neural Networks
boundaries have solidified. The essence of Artificial
Artificial Neural Networks, or ANN's, are reliable tools
Intelligence techniques in engineering problem solving is
for predicting experimental data. They could use their
learning by presenting real examples of input data
computational power in learning and generalization. They
(Example: Images captured on autonomous steering
are widely used for solving nonlinear problems.
systems) and outputs (Example: Steering angle, braking
Prediction and generalization performance depend on
and acceleration at autonomous driving systems)
network training. Generalization is related to ANN
presented to them in such a way that subtle functional
performance with inputs that were not used during their
relationships between data are captured, even if the
training. The inspiration for artificial neural networks is
underlying relationships are unknown or the physical
the human brain, which uses parallel processing networks
meaning is difficult to explain [8].
used to control complex relationships between input and
output variables [9]. In Figure 2 a timeline of the
In a linear regression statistical model, the function f can
development of deep learning and neural networks is
be obtained by changing the slope of the curve, tan (Φ)
shown in the upper panel and the lower panel numerous
and intercept β of the straight line of Figure 1 (a), such
machine learning algorithms.
that the errors between the actual outputs and the straight
line outputs are minimized.
Deep Learning Apply BP to DBN & other
NN improvemnts

AleNet ResNet WGAN

Shallow Multylayers NN Stagnated Widely VGGNet, DenseNet

Neural & BP Algorithm applied GAN e
Network GoogleNet

1940’s 1960’s 1960’s-1970’s 1980’s Final dos 1990’s 2006 2009 2012 2014 2015 2016 2017
1980’s
K-Means Bayesian Non-linear Deep Forest
network SVM &
clustering

Machine
Decision
Learning tree

Figure 2: Deep learning development timeline and

commonly used machine learning algorithms. Source:
Adapted from [10].
Figure 1: Linear regression versus artificial intelligence
In ANN, simple artificial nodes, known as neurons,
(AI) model. (a) Linear regression modeling; (b) Data-
processing elements, or units, are connected to build a
driven AI modeling [8].
network that mimics the biological neural network of
humans. Back-propagation ANN is a well-known type of
The same principle is used in AI models. Artificial
neural network, with multilayer perceptron architecture
Intelligence can form a simple linear regression model
with error backpropagation for supervised learning and is
from one input and one output, Figure 1 (b), and use
particularly powerful for nonlinear prediction [11].
available data to map between system inputs and
corresponding outputs using repeated display machine
learning examples of model inputs and outputs (Training) 2.3. General Considerations on Artificial
in order to find the Y = f (X) function that minimizes the Neural Networks
error between historical (actual) outputs and AI model
predicted outputs [8]. The attraction of Artificial Neural Networks comes from
the remarkable information processing characteristic of
As common examples of artificial intelligence use it is biological systems, such as nonlinearity, high parallelism,
possible to mention: robustness, fault tolerance, learning, ability to handle

E-ISSN: 2224-3410 81 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

inaccurate information and their ability to generalize [12]. analogous to the biological neuron shown in Figure 3. It
Its ability to learn and processing information classifies it establishes a mapping between input activities (Stimulus)
as a form of Artificial Intelligence [13]. The most notable and the output signal. Perceptron presents n biological
feature of this technology is that it can be applied to a neurons with various signals of intensity x and synaptic
wide variety of problems, many of which were extremely strength w feeding a neuron with a minimum required
complex or lacking in more sophisticated theoretical stimulus value of b and the equivalent artificial neuron
models [14]. Figure 3 illustrates a scheme of a biological system. The Artificial Neural Network and the Biological
neuron with its three major functional units: dendrite, cell Network learn by adjusting the magnitude of synapse
body, and axon. The cell body has a nucleus that contains weights or forces [15].
information about hereditary characteristics and a plasma
in which resides the molecular equipment used by the 2.4. Deep Learning and Machine Learning
neuron in the process of communication with other
neurons [15]. In a simple definition, machine learning or deep learning
refers to the use of an artificial neural network with
multiple layers of hidden nodes between output and
input, as shown in Figure 5, where deep architectures are
built on multiple levels on nonlinear operations [17].

Output Layer

Input Layer Hidden Layers

Figure 3: The biological neuron [16].

Figure 5: Artificial Neural Network with multiple layers.
Source: Adapted from [17].

As shown in Figure 5, the deep network has the same

architecture of a traditional neural network, but with a
greater number of hidden layers. The main difference
between a deep network and a traditional neural network
is the development of algorithms for deep architecture
training, which are faster and yield stronger results. Deep
learning includes representation learning algorithms that
transform raw characteristics into high-level abstractions
using a deep network composed of several hidden layers.
In other words, deep learning applies computational
approaches, which have nonlinear multiple
Figure 4: Signal interactions from n neurons and analogy transformations to train data representation through
to the sum of signals in a single-layer artificial neuron various levels of abstraction [17].
[15].

A schematic illustration of the signal transfer between

3. Failures in Artificial Intelligence
two neurons across the synapse is shown in Figure 4. An
impulse, in the form of an electrical signal, travels within 3.1. Adversarial Examples
the dendrite and through the cell body toward the
presynaptic membrane of the synapse. Once the Applications in deep learning-driven systems are steadily
membrane is reached, a chemical neurotransmitter is increasing in the real world. For example, information
released from the vesicles in amounts proportional to the technology and auto companies (Google, Tesla,
strength of the newly arrived signal. The neurotransmitter Mercedes-Benz and Uber) are testing autonomous cars,
diffuses into the synaptic breach toward the postsynaptic which require a fullness of deep learning techniques such
membrane and eventually into the dendrites of as object recognition, reinforcement learning and
neighboring neurons, thereby forcing them (depending on multimodal learning [18]. Opposing examples in
the minimum value required for stimulating the receiving conventional machine learning models have been
neuron) to generate a new electrical signal [15]. The discussed for decades. Machine learning based systems
system comprised of an artificial neuron and the inputs, with hand-made features are the primary targets such as
as shown in Figure 4, is called Perceptron and is

E-ISSN: 2224-3410 82 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

spam filters, intruder detection, biometric authentication,

fraud detection, and so on. The problem of Adversarial Image
Example was formulated as a match between the classified as
a Panda with
adversary and the classifier, both of which were cost 60% of
sensitive [18]. Attack and defense in opposing examples confidence.
have become an iterative game. Given the high
generalization capacity of deep neural networks, the
entire research community was surprised by finding that
these deep neural networks can be fooled by subtly Small
disturbing input data [19]. This type of data is sound adversarial
input data for machine learning models, but they are input.
intentionally designed by an attacker to cause the model
to fail. Figure 6 demonstrates the idea of what the
opposing examples look like. Starting with the image of a
panda, the attacker adds a small disturbance that has been
calculated to make the image look like a gibbon. Imperceptible
Opposing examples have the potential to be dangerous. modified
For example, attackers may target autonomous vehicles image,
classified as a
using stickers or paints to create an opposing stop signal Gibbon with
that the vehicle could interpret as a preference signal or 99% of
other traffic signal [20]. confidence.

Attackers can generate opposing commands against Figure 6: An opposing example constructed by modifying
automatic speech recognition models in speech- a panda's image in such a way that a machine learning
controllable systems such as Apple Siri, Amazon Alexa model thinks it is a gibbon [20].
and Microsoft Cortana [18]. The cause of these opposing
examples was a mystery, and speculative explanations
suggested that it was due to the extreme nonlinearity of 3.2. Soft Errors
deep neural networks, perhaps combined with the
insufficient average of the model and insufficient The ever-increasing miniaturization of semiconductors
regularization of the purely supervised learning problem. leads to major advances in mobile, cloud and network
computing. However, this has made electronic devices
A variety of intriguing properties of artificial neural less reliable and microprocessors more susceptible to
networks and related models have been demonstrated transient errors. These intermittent failures do not cause
[21]. But the consequences of errors increase permanent damage but may result in program execution
dramatically when technology companies start using deep by changing the transfer of stored signals or values.
learning algorithms in applications such as two-ton These transient faults are also called soft errors. As
machines moving on high-speed highways. A wrong technology continues to escalate, industry experts project
decision made by autonomous artificial intelligence can that the problem of soft errors will become increasingly
lead the car to collide with the guardrail, another vehicle important [23]. Deep learning neural network-based
or run over pedestrians or cyclists [22]. applications are widely used in high performance
computing systems and data centers.

While the performance of deep learning accelerators and

applications has been extensively studied, the
implications for reliability in their use are not well
understood. One of the biggest sources of unreliability in
modern systems are the soft errors. Such mild errors can
cause application failures and result in violations of
safety and reliability specifications. For example, in
autonomous cars, a slight error can lead to
misclassification of objects, resulting in a wrong action
taken by the car. In fault insertion experiments, many
cases were found where a truck could be misclassified
under a soft error, as shown in Figure 9.

E-ISSN: 2224-3410 83 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

Table 1: Summary of Characteristics of Artificial Neural

Networks

Characteristic Criteria MatLab

Training set =
Partition of data set 75%, RANPERM
Validation set
= 25%.
Figure 7: (Left) Fault Free Execution: A truck is Normalization Formula
identified by the deep neural network and then the brakes Net weigh
- INITNW
are applied. (Right) The truck is incorrectly identified as initialization
a bird and the brakes may not be applied [24]. Net learning ratio - TRAINGDX
Transfer Function - TANSIG
The deep neural network in the car should classify the Convergence
-
object ahead as a haul truck in a fault-free execution and Criteria
then apply the brakes in time to avoid a collision. Minimum error
0,001 -
However, due to a slight error in the deep neural network, aimed
the truck is misclassified as a bird, and the braking action Number of training
700 -
may not be applied in time to avoid a collision, especially cycle
when the car is operating at high speeds [24]. Training mode BT -
Number of hidden
1 -
3.3. Errors due the lack of appropriate layers
training Size of hidden layer 6 -
Net training mode - TRAINBR
The lack of enough data for appropriate neural network
training is associated with lower prediction accuracy or The figure 9 compares the actual and simulated values by
higher error. The section 4 will discuss a real case where the 111 samples model (25% of the original database)
the maximum and the minimum of the data, represented used in the artificial neural network validation tests. It is
by the tails of a histogram, presents a bigger error if noticed that the neural network better simulates the
compared with the average data, represented by the values close to the TS average, not being perfectly
central region of the histogram. accurate in the simulation of data that deviate much from
the average (larger difference between simulated and
4. Development of a mathematical measured points). This fact is due to the training process
model based on Artificial Intelligence being effective in the regions with the highest data
concentration [5].
In an artificial intelligence model developed for modeling
Comparison between the measured
the mechanical properties of hot rolled steel structural and the estimated steel variable
beams, a larger prediction error was found in regions with Measured
few occurrences of training data. The model was built Estimated

from three artificial neural networks of the Reverse

Normalized amplitude

Propagation type, using the MATLAB software. The

variable chosen for modeling is the Tensile Strength, TS,
one of the strength measures of materials. Figure 8 shows
the evolution of the Sum of Squared Errors (SSE)
calculated for the Neural Network training data
configured with 6 neurons to predict Tensile Strength.

Sample

Figure 9: Result of ANN Validation for TS prediction [5].

As the histogram in Figure 10 shows, the central region

concentrates most occurrences of learning data. In the
Figure 8: ANN training using MATLAB for TS tails, due to the absence of larger amount of data, as in
prediction (Oliveira, 2008). any normal distribution, the learning does not occur as in
the central region, which is evidenced by the biggest error
The table 1 summarizes all the criteria used to build the and smaller error of the simulated data, respectively. As
Artificial Neural Networks for the prediction of the seen in section 3.1, where autonomous vehicle accidents
tensile Strength. were discussed under rare environmental conditions, the
example shown in section 4 is similar. The tails of the

E-ISSN: 2224-3410 84 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

histogram correspond to the rare conditions in which area, however the Artificial Neural Networks stand out
there was insufficient neural network training to and are the basis of the so-called Deep Learning. These
modelling an important feature. Due to this fact, the error networks simulate the learning process of the human
found in this region is significantly higher when brain and learn through training with historical data.
compared to the central region of the histogram, as can be Technologies used in artificial intelligence tools have
seen in Figure 9. In this case, the estimated TS, been found to be subject to errors, known as adversary
represented by the blue curve, as a function of the examples and subtle errors, which are critical in certain
sample, presents greater error in regions coincident with uses where safety is a primary issue and lives may be
the highest normalized amplitudes (corresponding to the subject to decisions made by algorithms. An example is
maximum and minimum) of the measured TS, the advent of autonomous cars, subject to both
represented by the red curve. adversarial examples and subtle errors. The lack of
enough data on extreme situations are also critical for
training, prediction and subsequent action for uses where
safety is critical. A real case of building an artificial
intelligence for industrial application was discussed and
the possible measures to reduce the prediction error
obtained were pointed out. This article is expected to
contribute to the technical and professional growth of the
readers.

6. References
Figure 10: Histogram of the measured values of the [1] B. J. Copeland. (2019) Artificial intelligence.
Tensile Strength, [5]. https://www.britannica.com/technology/artificial-
intelligence;
It is proposed that, in the planning stage of the Artificial
Neural Networks model for the applications discussed, [2] Danny Y., & Dan T. 2016). Tesla driver dies in first
either in autonomous cars or in predictive models for fatal crash while using autopilot mode.
mechanical properties, it should be considered an https://www.theguardian.com/technology/2016/jun/30/tes
experimental design phase (Design of Experiments, la-autopilot-death-self-driving-car-elon-musk;
DOE), where the data related to the tails of the histogram
are intentionally generated and therefore would provide [3] Chen, A. (2018). IBM’s Watson gave unsafe
greater learning on specific and rare situations. For recommendations for treating cancer.
example, on the case of a recent car accident the https://www.theverge.com/2018/7/26/17619382/ibms-
autonomous driving system was not capable to take any watson-cancer-ai-healthcare-science;
action to avoid the collision with an overturned truck and,
also, ignored a pedestrian, at the same time [25]. Some [4] Vincent, J. (2016). Twitter taught Microsoft’s AI
extreme and possible situations, like an overturned truck chatbot to be a racist asshole in less than a day.
and similar objects lying down on the road, could be on https://www.theverge.com/2016/3/24/11297050/tay-
the training data to allow the car to learn that on microsoft-chatbot-racist;
situations like that, where the action of breaking is one of
the possible measures to avoid an accident. But it appears [5] Oliveira, A. P. Modelo de Previsão de propriedades
not to be the case on this specific example. For the mecânicas de perfis estruturais laminados a quente: uma
practical example discussed in section 4, simply design abordagem em redes neurais artificiais. (Dissertação,
the DOE such that the maximum and minimum regions Mestrado em Engenharia Metalúrgica e de Minas).
of the input variables have enough occurrences of Biblioteca Digital da Universidade Federal de Minas
training data. For this, the process must be modified in its Gerais, UFMG, 2008. http://hdl.handle.net/1843/MAPO-
origin, thus forcing the occurrence of these data [5]. 7RLKBJ;

[6] Lohr, Steve. (2018). Vencendo os limites da pesquisa

5. Conclusion no campo da inteligência artificial.
https://internacional.estadao.com.br/noticias/nytiw,vence
This paper discussed the main technologies used in
ndo-os-limites-da-pesquisa-no-campo-da-inteligencia-
Artificial Intelligence, their development history,
artificial,70002399992;
considerations about Artificial Neural Networks and the
failures arising from the training processes, hardware
[7] Hassabis, D., Kumaran, D., Summerfield, C.,
and lack of enough training data. It has been seen that
Botvinick, M. Neuroscience-Inspired artificial
these models learn as they are exposed to input and
intelligence. Neuron, 95, 2017, 245–258.
output data of any phenomenon that one wishes to have
http://dx.doi.org/10.1016/j.neuron.2017.06.011;
predictive ability. There are numerous technologies in the

E-ISSN: 2224-3410 85 Volume 17, 2020

 WSEAS TRANSACTIONS on ADVANCES in ENGINEERING EDUCATION Alisson Paulo De Oliveira,
DOI: 10.37394/232010.2020.17.10 Hugo Ferreira Tadeu Braga

[8] Shahin, M. A. State-of-the-art review of some defenses for deep learning. arXiv preprint
artificial intelligence applications in pile foundations. arXiv:1712.07107, 2018;
Geoscience Frontiers 7, 2014, 33-44.
http://dx.doi.org/10.1016/j.gsf.2014.10.002; [19] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer,
Michael K. Reiter. Accessorize to a crime: Real and
[9] Muhammad Shahbaz, Syed A. Taqvi, Adrian Chun stealthy attacks on state-of-the-art face recognition. In
Minh Loy, Abrar Inayat, Fahim Uddin, Awais Bokhari, Proceedings of the 23rd ACM SIGSAC Conference on
Salman Raza Naqvi. Artificial neural network approach Computer and Communications Security, October 2016.
for the steam gasification of palm oil waste using bottom DOI: http://dx.doi.org/10.1145/2976749.2978392;
ash and CaO. Renewable Energy 132, 2019, 243-254;
[20] Goodfellow, Ian; Papernot, Nicolas; Huang, Sandy;
[10] Cao, C.; Liu, F.; Tan, H.; Song, D.; Shu, W.; Li, W.; Duan, Yan; Abbeel, Pieter; Clark, Jack. (2017).
Zhou, Y.; Bo, X.; Xie, Z. Deep Learning and Its “Attacking Machine Learning with Adversarial
Applications in Biomedicine. Genomics, Proteomics & Examples.” OpenAI. https://blog.openai.com/adversarial-
Bioinformatics, 16, 2018, 17–32; example-research/;

[11] Chia-Yen, L., & Tsung-Lun, T. Data science [21] Ian J. Goodfellow, Jonathon Shlens, and Christian
framework for variable selection, metrology prediction, Szegedy. Explaining and harnessing adversarial
and T process control in TFT-LCD manufacturing. examples. 2015. arXiv:1412.6572;
Robotics and Computer Integrated Manufacturing, 55,
2018, 76–87. https://doi.org/10.1016/j.rcim.2018.07.013; [22] Hsu, J. (2017). A New Way to Find Bugs in Self-
Driving AI Could Save Lives.
[12] Tu, J. V. Advantages and Disadvantages of Using https://spectrum.ieee.org/tech-talk/robotics/artificial-
Artificial Neural Networks versus Logistic Regression for intelligence/better-bug-hunts-in-selfdriving-car-ai-could-
Predicting Medical Outcomes. Journal of Clinical save-lives;
Epidemiology, 49, 11, 1996, 1125-1231;
https://doi.org/10.1016/S0895-4356(96)00002-9; [23] Q. Shi, H. Omar, and O. Khan, “Exploiting the
tradeoff between program accuracy and soft-error
[13] Korczak, P., Dyja H., Labuda E. Using Neural resiliency overhead for machine learning workloads,”
Netwoks Models for Predicting Mechanical Properties CoRR, vol. abs/1707.02589, 2017. [Online]. Available:
after Plate Rolling Processes. Journal of Materials http://arxiv.org/abs/1707.02589;
Processing Technology, Poland, vol.80, n.81, 1998, 481-
486; [24] Li, G., Hari, S. K. S., Sullivan, M., Tsai, T.,
Pattabiraman, K., Emer, J., & Keckler, S. W. (2017).
[14] Dyja H., & Korczak P. The thermal-mechanical and Understanding error propagation in deep learning neural
microstructural model for the FEM simulation of hot network (DNN) accelerators and applications. Paper
plate rolling. Journal of Materials Processing presentend on: International Conference for High
Technology, 92-93, 1999 463-467. Performance Computing, Networking, Storage and
https://doi.org/10.1016/S0924-0136(99)00215-0; Analysis, Denver, Colorado.
https://doi.org/10.1145/3126908.3126964;
[15] I.A. Basheer, M. Hajmeer. Artificial Neural
Networks: Fundamentals, Computing, Design and [25] Templeton B. (2020). Tesla in Taiwan Crashes
Application, Journal of Microbiological Methods, vol.43, Directly into Overturned Truck, Ignores Pedestrian, With
2000, 3-31; Autopilot On.
https://www.forbes.com/sites/bradtempleton/2020/06/02/t
[16] Castrounis. (2016). Artificial Intelligence, Deep esla-in-taiwan-crashes-directly-into-overturned-truck-
Learning, and Neural Networks, Explained. ignores-pedestrian-with-autopilot-on/#5ad97d7758e5.
https://www.kdnuggets.com/2016/10/artificial-
intelligence-deep-learning-neural-networks-
explained.html;
Creative Commons Attribution License 4.0
[17] Milad Zafar Nezhad, Najibesadat Sadati, Kai Yang, (Attribution 4.0 International, CC BY 4.0)
Dongxiao Zhu. A Deep Active Survival Analysis
approach for precision treatment recommendations: This article is published under the terms of the Creative
Application of prostate cancer. Expert Systems with Commons Attribution License 4.0
Applications 115, 2018, 16–26; https://creativecommons.org/licenses/by/4.0/deed.en_US

[18] Xiaoyong Yuan, Pan He, Qile Zhu, Rajendra Rana

Bhat, Xiaolin Li. Adversarial examples: Attacks and

E-ISSN: 2224-3410 86 Volume 17, 2020