Научные исследования
и инновации в индустрии 4.0
I-
Республиканская научно
-
техническая конференция
(Ташкент,
13-15
марта
2022
года)
10
APPLICATION OF NEURAL NETWORKS
Gulyamov Saidakhror,
Academician, Doctor of economic sciences, Professor,
Institute for Retraining of Personnel and Statistical Research of the State
Statistics Committee of the Republic of Uzbekistan
Shermukhamedov Abbas,
Doctor of Physical and Mathematical Sciences, prof., Tashkent branch of the
Russian Economic University named after G.V. Plekhanov
Kholboev Bokhodir,
PhD, associate professor, Tashkent branch of the Russian Economic
University named after G.V. Plekhanov
Annotation:
Abstract: The article discusses neural networks that are
widely used in various fields, such as economics (prediction of stock market
indicators, prediction of financial time series), robotics (recognition of optical
and audio signals, self-learning), visualization of multidimensional data,
associative search for textual information, etc.
Neural networks are of interest to a fairly large number of specialists, for
example for computer scientists’ neural networks open up the field of new
methods for solving complex problems; physicists use neural networks to
model phenomena in statistical mechanics and to solve many other problems;
neurophysiologists can use neural networks to model and study brain
functions; psychologists have at their disposal a mechanism for testing
models of some of their psychological theory.
Keywords: perceptron, neural networks, input information, human
brain, neurobiology, electromagnetic activity.
The main area of research on artificial intelligence in the 1960s –
1980s. There were expert systems that were based on high-level modeling of
the thinking process (in particular, on the idea that our thinking process is
based on manipulating symbols).
However, it soon became clear that such systems, although they may
be useful in some areas; do not cover some key aspects of human
intelligence.
According, to one of the common points of view, the reason is that they
are not able to reproduce the structure of the brain, and in order to create
artificial intelligence, it is necessary to build a system with a similar
architecture. The human brain consists of a very large number (approximately
10,000,000,000) neurons connected by numerous connections (on average,
several thousand connections per neuron, however, this value can vary
greatly).
Научные исследования
и инновации в индустрии 4.0
I-
Республиканская научно
-
техническая конференция
(Ташкент,
13-15
марта
2022
года)
11
A neuron is a special cell capable of transmitting electrochemical
signals. A neuron has a branched structure of input channels of information
(dendrites), a nucleus and a branching output channel (axon). The axons of
such a cell are connected to the dendrites of other neurons in the cells using
synapses. When activated, a neuron sends an electrochemical signal along
its axon, and through synapses, this signal reaches other neurons, which can,
in turn, be activated. A neuron is activated when the total level of signals
arriving at its nucleus from dendrites exceeds a certain level (activation
threshold).
The intensity of the signal received by the neuron (and, consequently,
the possibility of its activation) strongly depends on the activity of synapses.
Each synapse is a gap (synaptic cleft) between an axon and a dendrite, and
special chemicals (neurotransmitters) transmit a signal through this gap. One
of the most respected researchers of neurosystems, Donald Hebb,
formulated the postulate that learning is primarily about changes in the
“strength” of synaptic connections. For example, in Pavlov’s classic
experiment, a bell rang each time just before feeding the dog, and the dog
quickly learned to associate the bell with food. This happened because the
synaptic connections between the parts of the cerebral cortex responsible for
hearing and the salivary glands increased, so that when the cerebral cortex
was excited by the sound of a bell, the dog began to salivate. In this way.
In practice, neural networks are used, as software products that run on
ordinary computers, or as specialized hardware and software systems [2].
Note, that in the first case, the built-in parallelism of neural network
algorithms is most often not used, since for many tasks of analyzing and
generalizing databases, special performance is not required – for them, the
performance of modern universal processors is quite enough. Such
applications use exclusively the ability of neural networks to learn and to
extract patterns hidden in large amounts of information. For the second group
of applications, usually associated with real-time signal processing, the
parallelism of neural computations is a critical factor.
The main tasks are solved by neural networks:
1. Distributed associative memory. Distributed memory means that the
weights of the connections of neurons have the status of information without
a specific association of a piece of information with a particular neuron.
Associative memory means that a neural network is able to output a complete
image from the part presented at the input.
2. Pattern recognition. Pattern recognition tasks require the ability to
simultaneously process a large amount of input information and produce a
categorical or generalized answer. For this, the neural network must have
internal parallelism.
3. Adaptive management.
Научные исследования
и инновации в индустрии 4.0
I-
Республиканская научно
-
техническая конференция
(Ташкент,
13-15
марта
2022
года)
12
4. Forecasting.
5. Expert systems.
6. Optimization (i.e., the search for the maximum of the functional in the
presence of restrictions on its parameters).
Currently, neural networks are widely used in various fields, such as
economics (predicting stock market indicators, predicting financial time
series), robotics (recognizing optical and sound signals, self-learning),
visualization of multidimensional data, associative search for textual
information, etc.
Neural networks are of interest for a large number of specialists:
For computer scientists, neural networks open up the field of new
methods for solving complex problems;
Physicists use neural networks to model phenomena in statistical
mechanics and to solve many other problems;
Neurophysiologists can use neural networks to model and study brain
functions; psychologists have at their disposal a mechanism for testing
models of some of their psychological theories;
Neural networks, other specialists (especially commercial and industrial
areas) may also be interested in neural networks for a variety of reasons,
primarily due to the new possibilities of forecasting and data visualization
achieved with their help. Learning Neural Network Learning is a fundamental
property of the brain.
In the context of neural networks, the learning process can be
considered as setting up the network architecture and connection weights
efficiently to perform some special tasks. Typically, a neural network must
adjust its link weights based on the available training set [1], and the network
performance improves as the weights are iteratively adjusted. The property of
neural networks to learn by example makes them more attractive in
comparison with systems, which work according to a rigidly defined set of
functioning rules formulated by experts. To design the learning process of a
neural network, first of all, it is necessary to have a model of the external
environment in which this neural network should function, i.e., to know the
information available to the network. This model defines the learning
paradigm. Secondly, it is necessary to understand how exactly the weights of
the network should be modified, that is, which learning rules govern the
tuning process.
Learning Algorithm refers to a procedure that uses learning rules to set
up weights. There are three paradigms for teaching neural networks: “with a
teacher”, “without a teacher” (self-learning) and mixed [1]. In the first case,
the neural network has the correct answers (the required outputs of the
network) for each input example, and the weights adjusted, so that the
network produces answers as close as possible to the known correct
answers.
Научные исследования
и инновации в индустрии 4.0
I-
Республиканская научно
-
техническая конференция
(Ташкент,
13-15
марта
2022
года)
13
The strengthened version of learning “with a teacher” assumes that only
a critical assessment of the correctness of the output of the neural network is
known, but not the correct values of the output themselves. Unsupervised
learning does not require knowing the correct answers for each training
sample. In this case, only the internal structure of the data or the correlations
between the samples in the data system is revealed, which allows the
samples to be categorized. In blended learning, some of the weights are
determined through supervised learning, while the rest of the weights formed
through self-learning.
Neural network learning theory considers three fundamental properties
associated with learning for example capacity, sample complexity, and
computational complexity. In this case, the capacity is understood as how
many samples the network can remember and what functions and decision-
making boundaries can be formed on it. The complexity of the samples
determines the number of training examples needed to achieve the
generalizability of the network. Too few of these examples can cause the
network to “overfit” when it performs well on training set examples, but
performs poorly on test cases subject to the same statistical distribution [3].
There are four main types of learning rules: error correction, Boltzmann
machine, Hebb’s rule, and competition learning. Error correction rule. In
supervised learning, the desired output d is given for each input example, but
the actual output of the network y may not coincide with the desired one. The
principle of error correction during learning is to use a difference signal (d – y)
to modify the weights to gradually reduce the error. Such training is
performed only when the neural network is wrong.
Moreover, there are various modifications of this learning algorithm,
which are not discussed in detail here. Neural networks 125 Boltzmann
training. It is a stochastic learning rule that follows the principles of
information theory and thermodynamic principles.
The goal of Boltzmann’s training is to adjust the weighting coefficients in
such a way that the states of the neurons of the outer layer satisfy the desired
probability distribution. Boltzmann training can be viewed as a special case of
error correction, in which the error is understood as the discrepancy between
the state correlations in two modes. Hebb’s rule.
The oldest teaching rule is Hebb’s teaching postulate. Hebb relied on
the following neurophysiological observations: if neurons on both sides of the
synapse are fired simultaneously and regularly, then the strength of the
synaptic connection increases. An important feature of this rule is that the
change in synaptic weight here depends only on the activity of neurons that
are connected by a given synapse. Competition training. Unlike Hebb
learning, in which multiple output neurons can be fired simultaneously, in
competitive learning, the output neurons compete with each other for firing.
This phenomenon is known as the “winner-take-all” rule. Similar learning
Научные исследования
и инновации в индустрии 4.0
I-
Республиканская научно
-
техническая конференция
(Ташкент,
13-15
марта
2022
года)
14
takes place in biological neural networks. Learning through competition
allows you to cluster your inputs: similar input examples are grouped by the
network according to correlations and represented by a single element. When
learning by the method of competition, only the weights of the “winning”
neuron are modified. The effect of this rule is achieved, due to such a change
in the sample stored in the network (the vector of connection weights of the
“winning” neuron), in which it becomes a little closer to the input example.
REFERENCES:
1. Galushkin A.I. Theory of neural networks. – Moscow: Publishing
company of the editorial office of the magazine “Radiotekhnika”, 2000.
2. Neural networks in automation systems / V.I. Arkhangelsky,
I.N. Bogaenko, G.G. Grabovsky et al. – Kiev: Technics, 1999.
3. Terekhov S.A. Lectures on the theory and applications of artificial
neural networks: http://alife.narod.ru/lectures/neural/ Neu_index.htm.
ФОРМИРОВАНИЕ ЭФФЕКТИВНОЙ ОБРАЗОВАТЕЛЬНОЙ
СРЕДЫ
В УСЛОВИЯХ РАЗВИТИЯ ИНДУСТРИИ 4.0.
Зокирова Нодира Каландаровна,
д.э.н., профессор, зав. каф. «Экономика труда и
управление».редседатель Совета по науке и инновациям
Ташкентского филиала РЭУ им. Г.В. Плеханова
Новый индустриальный уклад, характеризующийся стремительной
диффузией инновационных и цифровых технологий в экономике,
становится локомотивом повышения качества образовательной
среды.
Например, выдвигая первостепенной задачей, достижение
адекватного
уровня,
параметров
эффективности
и
конкурентоспособности образовательных организаций. Преимущества,
четвертой промышленной революции проявляются, не только в
массовом внедрении информационных технологий, но и в создании
роботизированных систем и автоматизации производственных
процессов. А также, в обеспечении ускорения интеграционных
процессов в образовательной среде, с учетом, расширения
возможностей использования современных
производственных систем в
качестве платформ для приобретения практических навыков и
компетенций обучающимися на системной основе. Это несомненно,
снижает риски ошибок в выборе профессии и повышает качество
подготовки специалистов для индустрии 4.0.