2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

A Brief Review on Artificial Neural Network:

Network Structures and Applications
2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS) | 979-8-3503-9737-6/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICACCS57279.2023.10112753

1
Nilanjan Saha, 2Aleena Swetapadma,3 Mahadev Mondal.
School of computer engineering Kalinga Institute of Industrial
Technology Deemed to be University Bhubaneshwar, India
1
sahanilanjan807@gmail.com,
2
aleena.swetapadmafcs@kiit.ac.in, 3mondalmahadev04@gmail.com

Abstract—A quick overview of artificial neural the US. Neural Networks for Computing, which has since
networks (ANN) and its real-life applications are provided in become an annual conference, was first held by the
this publication. Different type of Neural Networks and their American Instituteof Physics in 1985.
structures are also been discussed here. Neurons in neural
networks are interconnected with other neurons to mimic By Schmidhuber & Hochreiter, a recurrent neural
how the intricate human brain functions. Artificial network framework called LSTM was proposed in 1997.
intelligence benefits from this because neural networks can Gradient-Based Learning Applied to Document
develop over time without human intervention once they are Recognition was published by Yann LeCun in 1998. To get
given a set of rules or guidelines, whereas AI can start from us to this point, additional actions were taken. Numerous
scratch, create its algorithm, and learn the rule with the aid applications of neural networks exist, some of which are
of machine learning techniques based on the rules. AI does mentioned later in this paper. The underlying premise of
not require the baseline since it can learn from observations, neural networks is that if something functions in nature, it
whereas neural networks learn through human feedback must also function in computers.
from the initial baselines.
The goal of simulating the human brain in a neural
Keywords—ANN, Back-propagation Neural Network, network is to produce a machine learning model that can
Kohonen Network, Grossberg Network, Long Short-Term achieve tasks which are challenging for computers when
MemoryNetworks, CNN. we using conventional conclusion. The human brain is
suitable of filtering huge information, admit designs, and
I. INTRODUCTION creating conclusions based on insufficient or unclear data.
In 1943, article on the potential functioning of
AI can have advantage from neural networks in many
neurons byWalter Pitts & Warren McCulloch constructed
ways like it helps Neural networks to conclude based on
a straightforwardnetwork utilizing electrically powered
huge datasets. It can master how to understand styles
circuits. Donald Hebb furthered understanding ofneural
and relationships in records. It can also beimproved to carry
neurons in 1942 when Organization of Behavior was
out large estimation rapidly and correctly. It can modify to
published. It made the point that each time brain pathways
new and transform statistics inputs.
are engaged, they become stronger. In 1950, the initial
simulation of neural network was done under supervision II. CONCEPTS ASSOCIATED
of Nathanial Rochester who was from the IBM
Research Laboratories.Scientific study by the Dartmouth The basis of deep learning techniques is a subclass of
College on computing furnish lift to each computing & machine learning known as neural networks, sometimes
neural networks. known as artificial neural networks (ANNs) or simulated
neural networks (SNNs). Their structure and phrasing are
John,a mathematician steered emulating modeled mirrored the mortal(human) brain, imitating how
straightforward nerve cell actions by exploitation the communications of the real neurons happens. A node
telegraphy transfers /voids(vacuum) tubes in the year of layer, which includes an intake level(input layer), one or
1957. In 1958, Frank Rosenblatt began his project on more concealed layers, and an product level(output
Perceptron. It is the oldest Neural Networks,it is still in layer),this all makes up a neural network. Every node is
operate able situation.The primary Neural Network that connected to each others and has a weight and threshold
can be used in true world problems, were suggested by value. Any node those outcome exceeds the
Bernard Widrow and Marcian Hoff, from Stanford. They predetermined threshold value is triggered and begins
made the models that was refereedas Adaline & Madaline, relaying data to the network's higher layer. Consider a
in 1959. In 1981, advancement on Neural Network network where each node consists solely of input
analysis interrupted because of unsuccessful claims. John information(data), weights, a bias, & an outcome. Fig.1
Hopfield’s work was formulate and attractive and his illustrates the general form of equations used in developing
strategy was to make helpful machines , utilizing Neural a neural network. There are two types of neuron (in the
Network. In 1982 he published a journal to the National basis of number of inputs), those are single input neuron
Academy of Sciences. and multiple input neuron.
At the US Japan mutual convention on
Cooperative/Competitive Neural Networks in 1982, Japan
revealed their Fifth-Generation effort., the US became
concerned about falling behind and renewed financing in
979-8-3503-9737-6/23/$31.00 ©2023 IEEE

1974
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

data that has neither classed nor labeled and permitting the
algorithmic program to operate on that data in absence of
supervision. Lacking any previous data learning, the
device's goal in this event is to categories unorganized
information in accordance with correspondences,
paradigms, and dissimilarities.
5) Reinforcement learning: It is the machine
learning models that have been trained to make a series of
decisions, is described in section.The agent captures the
capacity to carry out a task in a situation that could be
challenging and unclear.
Artificial intelligence disputes a game like situation in
the course of reinforcement learning. In order to solving
the game(problem), the device exploits trials and errors.
Fig. 1. A simple neural network. Then the Artificial Intelligence is awarded or penalized.
A. Different Parameters Used to Design The Learning is the operation of determining what the
Neural Network AreGiven Below. estimates of the weights suppose to be. It is usually
accomplished through comparisons of the outcome of the
components with what this must be. Normally the Widrow
Hoff Delta rule is used to adjust weights. Such neurons are
capable of carrying out basic tasks in a single layer
network[1].
III. DIFFERENT TYPES OF NEURAL NETWORK
A. Perceptron Network
The McCulloch and Pitt’s neuron extension serves as
the primary component in these networks. The element's
output is a operator of the weighted sum total of all its
intakes. Occasionally, the result value is applied
immediately, or the sigmoid function or a threshold can
B. These two techniques can be used to minimize the process the output. The weighted aggregate must exceed a
cost function as discussed below: specific threshold in order for an element to be considered
1) Back propagation: In this process, data takes to have "fired". The exact output value is needed for other
entry in the input layer and travels to the output layer networks. Fig.2 shows a simple perceptron network.
through the network. The outcome will subsequently be
The process of determining what the weights' values
equal to the predicted output according to the cost
should be is known as learning. This is frequently
function. If the cost function's value in higher side then,
accomplished by comparing the element's output to what
the data is returned, and the neural network learns to reduce
it should be and then correctly modifying the weights.
the cost function's value by adjusting the weights.
The Widrow-Hoff Delta rule is typically applied to
2) Forward Propagation: Information (data) weight adjustments. Such neurons are capable of carrying
penetrates the input layer and propagates forward through out basic tasks in a single layer network.
network to the output value. The user compares the value
In the learning phase, the network is frequently
to the anticipated outcomes. The following stage involves
introduced to the training set's information, including the
computing errors and sending information backwards.
necessary intakes and outcomes on behalf of every
Now, the user can update the weights and train the neural
element of set, and load(weights) are modified as
network. The organized data allows the user to change
necessary using the generalized delta rule by propagating
weights simultaneously.
errors between the desired and actual outputs back through
Neural networks can be broadly classified based on the network. Since the entire set of information must be
the type of learning it used such as supervised learning, given hundredth times, it can take some time. This is a
unsupervised learning and reinforcement learning. serious drawback of the approach.
3) Supervised learning: In this method, the machine
learns under supervision . It includes a model that can
make predictions using labeled datasets. A datasets that
has been labeled means that you already know the
intended A datasets that has been labeled means that you
already know the intended A datasets that has been
labeled means that you already know the intended
4) Unsupervised learning: often known as machine
learning, is the process of learning a device utilizing the Fig. 2. Perceptron network.

1975
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

B. Feed-Forward Neural Network respectively, and A, B, and C denote the leak rate or the
test cycle time for a pressure Decay test and the neuron's
This is the most general type of Neural Network, highest and minimal activating layers, respectively. When
where incominginformationflows in only one direction, the system is at equilibrium, the activation n achieves the
through artificial neural network , and out by way of value n=BE- CI/A+E+I [3].
output node. A gateway to large neural networks that
route input data through a number of artificial neuronal E. Deep Neural Network:
levels. It’s every node is attached to every single neuron, The main difference between Neural Network and
therefore it is completely connected neural network. Deep Neural Network(DNN) is the number of hidden
layers. DNN has minimum 3 hidden layers where neural
network has one or two hidden layers.Different type of
DNN is discussed here:

1) Convolutional Neural Network (CNN):CNN is

different from traditional neural networks. CNN is mainly
used for image, audio and video analysis. Traditional
neural network uses totally connected layers in the other
hand CNN uses set of filters to analyze the sample data. A
CNN uses a modified version of multi layer perceptron.
There may be one or more convolutional layers in a CNN.
Either all of these layers can be joined together or they can
Fig. 3. Feed forward network. be combined.
C. Kohonen Network 2) Recurrent Neural Networks (RNN): Recurrent neural
networks are fed back to the input in order to better
Teuvo Kohonen proposed an alternative network
forecast how the layer will perform.
structure and learning method. These networks learn in an
unsupervised manner, which is more akin to how people 3) Long Short Term Memory (LSTM) Networks: LSTM
learn in some ways. Each node in Kohonen networks has a is a kind of Radial basis network that assists exceptional
number of values connected to it, one value for each units apart from some conventional units. one "memory cell"
network input. and each node's input is connected to it. can be initiate in Long Short Term Memory units that is
Each node's related values are compared to the input when able of containing information in memories for prolonged
the system is in operation. and the node that matches the period.
input the most is the one that fires. The system needs to
learn which patterns to fire and how to spot similar 4) Sequence to sequence models : Sequence to sequence
patterns.The next method is used to accomplish this. The models are those where the input and output are processed
node that most closely resembles the input is chosen in byan encoder and a decoder, respectively.
learn mode. After then, its weights are modified. to lessen 5) Modular Neural Network: A commutable(modular)
the difference between the inputs and the weights. This Neural Network consists innumerable distinct networks,
enables the node to more nearly match the input. The those networks carry out a certain task independently [4].
nodes that are close to this node, or those that are
connected to it, also have their weights altered so that IV. APPLICATIONS OF NEURAL NETWORK
similar patterns can be recalled; however, the amount by
A. Application In the Field Of Medicine
which they are adjusted depends on the deference between
each leaf node and the node that have been fired. Other 1) Breast Cancer: Breast cancer is a fatal disease
nodes are altered by M*δ if the main node is altered by an Among women, it has a high death rate. Ongoing work is
amount δ, where M is chosen by the "Mexican hat" being done to develop more effective methods for rapid
function [2]. and precise diagnosis. According to traditional
D. Grossberg Network approaches, doctors must inspect breast lesions to identify
distinct cancer stages. Such approaches are frequently
It was initiated by Stephen Grossberg. So this is a self
assembling, combative network built on constant period. This time-consuming and ineffective. Therefore, there is a need
network was created by neuroscientist and biomedical for effective techniques that accurately and without human
engineer Grossberg and is based on the human visual intervention diagnosis malignant cells. In order to
system. establish imaging biomarkers through mammography
analysis and on the basis of artificial intelligence
Among the Grossberg’s Neural Network
technology, image processing techniques were applied.
representations, the shunting model is attributed to
leaking(leaky) consolidator and is characterized with a The goal was to identify breast cancer in its early stages to
derived function(equation). support diagnosis and priorities high-risk patients. A
computer-based diagnosing procedure is being produced
using bio-markers and skilled neural networking. The
outcomes demonstrated that general regression ANN is a
here n=n(t) denotes a neuron's level of activation, reliable and promising method for finding breast cancer
E=E(t) and I=I(t) denote its excitatory and inhibitory inputs, [5, 6, 7].

1976
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

2) Analysis of ECGs: Clinical practise makes networks, shows how to classify a vehicle into one of
extensive use of computational ECG analyzers. For the several types after detection (Belgaroui and Blosseville,
analysis of ECG data, currently existing methods are 1993). Thought of a technique preserving multiple Neural
mostly established on rules-based statistical and networks, each one carrying out a sub-activity. Neural
mathematical methods. Numerous initiatives have been networks have been utilized in the complication of
made for using Neural Networks for increasing distinctive automated number plate scanning (Margarita,1990). This
precision and obtain perfect process even in the facing of task can be simply added with detecting systems.[12].
confounding variables. Examination of long-standing
2) Clinical Traffic pattern analysis: Trafficking
ECG records (by using Holter-monitor) needs automatic
networks with arrays of introductory loops or other
acknowledgment of sporadic occurrences; human
comparable detectors are productive origins of
evaluation of roughly 90.000 ECG-compounds daily is a
information on variables like speed, volume etc of
laborious and long process. A neural network that has
passing vehicles. This datasets, specifically when
received the proper practice can identify specific illnesses
gathered from multiple different geographical locations,
with up to 99.99% sensitivity[8, 9].
are highly complex to examine because of the spatial and
3) Intensive care unit : In an intensive care unit, temporal random relationships that determine the behavior
clinical parameters are currently assessed individually, of transportation systems. As such, several groups are
with a focus on the usage of alarm signals. A good alarm engaged in task with the common goal of utilizing Neural
system would be one that can assess simultaneous Networks to find-out similarities in this information, and
changes in physical, chemical, and thermodynamical there exists broad applications in this field. Neural
factors. For such a system, conventional technology would Networks proved to be useful for congestion diagnosis
have to presuppose complete clinical relationship through learning the Neural Network to categorize
knowledge. The capacity for learning and adaptability of downtown transit networks within one of two the
neural networks could be anticipated to be used as these counties(overloaded or not) (Kirby et al., 1993). The
relationships are only partially understood. In the practise primary flaw of this project are the aforementioned
of anaesthesia, doctors must watch a rising number of portability difficulties. This is because network was
displays and analyse an rising amount of indications. For instructed on a very precised information set
this volume of information to be processed effectively, at determined through native experts. A later study used a
a high degree of systematization, and with prompt review similar approach (Hua and Faghri,1993), however
of their input, doctors with experience and expertise are extended the count of overload classifications. The study
required. In an emergency, neural networks can help uses adjustable resonance hypothesis.[13,14].
doctors by launching the essential responses (inside 17
3) Clinical Freight operations: A major issue in
seconds contrary of 45 seconds mean by doctor) [10, 11].
shipment operations, where efforts created to apply
B. Application In Transport Neural Networks, is related to optimizing routing
1) Clinical Vehicle detection/classification: Starting networks and scheduling. A major problem that seems to
with the objective of a expressway with sensor arrays, a lot have been experienced by all researchers is the
of work carried out to abstract the maximal number of data parameterisation on the issue. This occurs since the issue
from the generated indications. A standard derived is effectively nonlinear. A potential explanation to this,
amount is the group of moving vehicles. Components is the usage of individual- organizing classification
such as wheelbase, count of axles, and weights. Basis systems. Various optimization methods have been
function networks been implemented to this method with utilized relying on the classification. Alternative method
few achievement (Mead et al,1994). Though statistical is to search alternative coding methods in hopes of finding
outcomes came out unremarkable, industrial organizations one that extracts key characteristics.One more
utilizing the arithmetical approach carried out even Alternative method was to utilize the Boltzmann
terrible. Of course, much extended testing, probably with Machine. This type of network, unlike other paradigms,
equipment from different industrialists, is required to get was specially intended for optimization problems.
the full view.A more determined concept would be to unluckily the outcomes were a bit unsatisfying.(Potuin
abandon conservative vehicle sensing techniques plus use and Shen, 1991) Techniques with complex pre
powerful image processing techniques is to use a video processing appear to be more promising.[15].
camera. In combination with Two of his studies utilizing C. Application In Rainfall Forecasting
neural networks that elements mutually reported in this
The purpose upon this is to predict the amount of
area. The primary issues the detecting vehicles passing precipitation one step ahead. In developing the proposed
through a motion camera( Bullock et al., 1992-93). precipitation prediction model, the precipitation operation
Although their performance under ideal conditions is that was considered to be a Markov techniques. This assets
equivalent to traditional image manipulation facilities, in order that this precipitation rate at a particular
neural networks have been reported to be more flexible to situation in outer- space and timing is the operation of a
modifications in outside factors for instance, shadows and defined series of previous comprehension. This model
camera position. A second study, again using neural configuration might be stated as: Here X(t + 1) + e (t),X

1977
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

(t) = g (X(t), X(t – l), X(t – 2, . . , X(t – k) + 1)) represents the implementation of hardware is reliant.
a vector of rain values x1t, x2t, . . , xNt at ‘N’ separate
sites,here t is time. g( ), it is the non-linear plotting 2) Mysterious Network Behavior: This is the main
operation estimated utilizing Artificial Neural Network, problem with ANNs. When ANN creates a probing
e(t) is the plotting flaw and k is the (unidentified) solution, it doesn't tell you why or how. This makes the
count of the precipitation realizations, contributing network less trustworthy.
precipitation for the following duration stage. k generally 3) Determining the correct system arrangement:
indicate the network delay. When the value of k is equal There are no definite requirements for deciding the
to 1 here, the network has a delay of 1 because the arrangement of ANN. Appropriate system arrangement is
precipitation in the following time step is only related to reached with knowledge, trial and mistake.
the precipitation that is currently occurring. Developing
an Artificial Neural Network for precipitation foretelling 4) Difficult to display issue to system: Artificial
includes the essential deliberations as mentioned below: Neural Network can manage numerical data. The problem
I)Choose an proper Artificial Neural Network to illustrate needs to be converted to numbers ahead of being presented
the Markov method II) Evaluate ANN delay. i.e., into the Artificial Neural Network. The show mechanism
Determine how many historical precipitation values to decided herein immediately affects the execution of the
include as input III)Determine the optimal ANN system.
complexity according the issue that is, resolute the count 5) Duration of the network is unknown:The
of total masked layers & the count of nodes in the masked system is reduction to a certain value probe failure means
layers [16]. practice is complete. The value does not deliver perfect
D. Application in stock market conclusions [18].
Neural networksanalyze price dataand uncover VI. CONCLUSION
opportunities. Neural networks allow traders to make
In terms of number and variety of applications, it is
trading decisions on the basison comprehensively
clear that the utilization of neural networks will increase
inspecting the information, which is not always true when
in future. Neural networking techniques require a plenty
employing conventional techniques for technological
of effort, feeding information on to the device, monitoring
inspection. For the profound and thoughtful broker,
performance, adjusting procedures, adding relations,
Artificial Neural Networks are a next generation (modern)
changing requirements, etc. unless the system realize the
equipment, which has a with terrific scope, capable of
required result. These required outcomes are statistic in
detecting elusive non linear dependencies & models that
quality. Networks are not at all times correct. Considering
further technical analysis system cannot reveal.[17].
this logic, Neural Networks are used in application
V. ADVANTAGES AND LIMITATION wherein people are not constantly correct. Neural
networks may nowadays examine task, optimize control
A. Advantages of Artificial Neural Network methods, decline credit cards, permit loans, market
Storing information over networks: Information like outlook and select shares. But the upcoming promises
traditional programming is stored overnetworks instead of better. Neural networks require quicker hardware.
databases. The loss of some information from one Integrating neural networks with fuzzy logic has
location does not hamper the functioning of the network. already yielded good results as a more effective and
efficient solution to problems that humans can better
1) Ability to handle partial information: Following
manage. The futur of neural networks is bright, and
Artificial Neural Network learning, Information can
current research seems to be on the right track
generate result even with unfinished knowledge.
The functioning degradation in this relies on value of the towards the ultimate goal of all artificial intelligence like
lackingdata. developing humanoid robots that can act and think like
humans.
2) Error Tolerant: Corruptness of single or multiple
cells of the Artificial Neural Network will not stop RERENCES
production of results. This characteristic improves [1] W.T. Illingworth, “Beginner's guide to neural networks",
network resilience. IEEE Aerospace and Electronic Systems Magazine, vol. 4,
1989.
3) Gradual Degradation: The system delay down [2] J.M.Bishop RJ. Mitchell, “Neural networks-an-introduction”.
very slowly and the experiences relative deterioration. [3] https://www.wikiwand.com/en/Grossberg_network
System issues don'terode quickly. [4] https://www.mygreatlearning.com/blog/types-of-neural-networks/
4) Machine Learning Capabilities: An artificial neural [5] O.O. Aalen, “Heterogeneity in survival analysis”, Statistics in
Medicine, vol. 7, pp. 1121–1137, 1988.
network learns events and annotates similar events to
[6] Abrahamowicz, M., Ciampi, A. and Ramsay, J. O.,
make decisions. “Nonparametric density estimation for censored survival data:
5) Parallel Processing Power: ANN have the regression-spline approach”, Canadian Journal of Statistics,
vol. 20, pp. 171–185, 1992.
mathematicalstrength to execute several work at once.
[7] Abrahamowicz, M., MacKenzie, T. and Esdaile, J. M, “Time-
B. Limitations of Artificial Neural Network dependent hazard ratio: modeling and hypothesis testing with
application in lupus nephritis”, 1996.
1) Hardware dependency: According to its structure, [8] Carroll TO, Ved H, Reilly D: A neural network for ECG
ANN need computers with parallel computing power. So, analysis. IJCNN Proceedings II: 575, 1989.

1978
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.
 2023 9th International Conference on Advanced Computing and Communication Systems (ICACCS)

[9] Edenbrandt L, Devine B, Macfarlane PW: Neural networks for

classification of ECG ST-T segments [see comments]. J
Electrocardiol, vol. 25(3), pp. 167-73, 1992.
[10] Mylrea KC, Orr JA, Westenskow DR: Integration of
monitoring intelligent alarm anaesthesia: neural networks -
can they help? J Clin Monit, vol. 9(1), pp. 31-7, 1993.
[11] Westenskow DR, Orr JA, Simon FH et al: Intelligent alarm
reduce anaesthesiologist’s response time to critical faults.
Anaesthesiology, vol. 77(6), pp. 1074-9, 1992.
[12] Belgaroui B. and Blossville J. M. (1993) A road traffic
application of neural techniques. Recherche Transports
Security, English IssueNo. 9,pp. 53-65.
[13] Kirby H. R.. Boyle R. D. and Dougherty M. S. (1993)
Recognition of road traffic patterns using neural networks.
Conf. on Informing Technologies for Construction . Civil
Engineering and Transport, Brunel.
[14] Hua J. and Faghri A. (1993a) Traffic mark classification
using artificial neural networks. Proc. Pacific Rim Conf.
Seattle.
[15] Potuin J.-Y. and Shen Y. (1991) A neural network approach
to the vehicle dispatching problem. IEEE Inf. Conf. on
Neural Networks,Singapore.
[16] K.C. LUK, J.E. BALL, A. SHARMA, “An Application of
Artificial Neural Networks for Rainfall Forecasting”, Water
Research Laboratory School of Civil and Environmental
Engineering The University of New South Wales, Australia.
[17] https://www.investopedia.com/articles/trading/06/neuralnetwo
rks.as p
[18] Maad M. Mijwel, “Artificial Neural Networks Advantages
and Disadvantages” Computer science, college of science,
University of Baghdad Baghdad, January 2018.

1979
Authorized licensed use limited to: FOM University of Applied Sciences. Downloaded on May 13,2023 at 15:27:24 UTC from IEEE Xplore. Restrictions apply.