Main Approaches and Ethical Considerations in Machine Learning

8.1 Main Approaches to Creating AI

8.1.1 Symbolic Programming and Expert Systems

Symbolic programming and expert systems stand out because of their unique approach to modeling human thinking and solving problems. This approach is based on the principle that knowledge and reasoning processes can be represented through symbols and a system of rules that operate on these symbols, thereby imitating the human capacity for analysis and inference.

At the heart of symbolic programming lies the idea that any knowledge or thought process can be decomposed into fundamental elements—symbols—and that further manipulation of these symbols makes it possible to solve various problems. For example, in mathematics, numbers and operations on them (addition, subtraction, and so on) can be regarded as symbols and rules for transforming them, which makes it possible to solve mathematical problems. In a similar way, symbolic programming in AI uses symbols to represent concepts and objects from the real world, while rules determine how these symbols can be combined or transformed in order to solve more complex tasks.

Expert systems are an application of symbolic programming in which they are used to imitate the decision-making process of a human expert in a specific field. They are based on a knowledge base consisting of factual data (for example, symptoms of a disease) and heuristic rules (for example, if symptom A and symptom B are present, then the possible diagnosis is C). Thus, when the system is given a task (for example, diagnosing a disease), it uses its knowledge base to analyze the information provided and generate a conclusion, imitating the human reasoning process.

To illustrate this, imagine an expert system developed to diagnose automobile faults. Its knowledge base contains information about which symptoms indicate specific problems with a car. For example, if the car does not start and a clicking sound is heard when attempting to start it, the system may use a rule stating that these symptoms indicate a possible starter failure. In this way, the expert system analyzes the input data (symptoms), compares them with the information in its knowledge base, and suggests the most likely cause of the problem, much as an experienced mechanic would do.

This approach makes it possible to create highly specialized AI systems capable of solving complex tasks in narrow domains where deep subject knowledge and specific expertise are required. It opens broad opportunities for automating decision-making processes and providing expert support in various fields, from medicine to technical maintenance.

Despite their significant advantages and contributions to the development of artificial intelligence, symbolic programming and expert systems face a number of drawbacks that limit their application and effectiveness in some areas.

Limited knowledge base: One of the key limitations of expert systems is their dependence on a predefined and carefully structured knowledge base. This means that the system can consider only those scenarios and solve only those tasks for which knowledge has been explicitly entered into the base. For example, an expert system created to diagnose automobile faults may fail to account for a new type of breakdown or nonstandard vehicle behavior if this information has not been included in its knowledge base.

Difficulty of maintenance and updating: Continuous development and changes in the subject domain require regular updates to the expert system’s knowledge base. This can be an extremely labor-intensive and costly process, especially in rapidly developing fields such as medicine or information technology. For example, in medicine, new studies, drugs, and treatment methods are constantly appearing, and an expert system used for diagnosis or treatment recommendations must be updated regularly in order to remain relevant.

Handling uncertainty and ambiguity: Many real-world problems are characterized by uncertainty and ambiguity, which symbolic systems often cannot handle effectively. Expert systems usually work with clearly defined rules and conditions, but they may have difficulty interpreting vague, incomplete, or contradictory data. For example, in clinical diagnosis a patient may present symptoms that overlap across several possible diseases, and the expert system may not be able to account for all the nuances of the clinical context and the physician’s personal experience.

Lack of flexibility and adaptability: Traditional expert systems do not have the ability to learn or adapt on the basis of new experience or data without human intervention. Unlike machine-learning systems, which can independently “learn” from new data and improve their performance over time, expert systems are limited by predefined knowledge and rules. This makes them less effective in dynamic or complex environments where conditions and requirements can change quickly.

8.1.2 The Machine-Learning-Based Approach

With the development of computer technologies and scientific research, machine learning has become an integral part of modern artificial intelligence, opening new horizons in the creation of autonomous systems capable of self-learning, adaptation, and improvement. This approach goes beyond traditional programming, where the behavior of a machine is rigidly set by a developer, and seeks to make machines capable of independently “learning” from the data with which they interact.

At the center of machine learning is the idea that machines can learn from experience, analyze data, and make decisions with minimal human intervention. This process begins with the collection and analysis of large volumes of data; on the basis of this data, the machine, using special algorithms, builds a model capable of making predictions or decisions under new conditions. Thus, unlike traditional programming, where the logic of a program is completely specified by the developer, in machine learning the machine itself “discovers” this logic by analyzing the data provided to it.

Machine-learning methods are classified according to the type of learning. Supervised learning, for example, requires a training dataset with explicitly specified correct answers (labels), allowing the machine to learn by matching input data with the corresponding outputs. This is like teaching a child to recognize fruit by showing an apple and saying that it is an apple; in this way the machine learns to recognize and classify data. In unsupervised learning, there are no labels, and the machine tries to find structure or patterns in the data on its own. This can be compared to a child trying independently to discover what different fruits have in common, without explicit instructions from an adult.

This approach allows machines to adapt to changes in data or in the environment while continuing to function effectively even under new or unforeseen conditions. This makes machine learning a powerful tool in areas such as pattern recognition, natural language processing, medical diagnostics, and many others, where systems are capable not only of following specified instructions but also of “learning” from acquired experience, improving their skills and increasing overall effectiveness.

This approach offers opportunities in various fields, but like any technology, it has both advantages and disadvantages.

Advantages of machine learning:

Adaptability: the ability to adapt to new changes in data, which makes such systems ideal for dynamically changing environments and requirements. - Automation of decision-making processes: the ability to automate complex decision-making processes by analyzing large volumes of data and providing predictions or recommendations without direct human participation. - Processing and analysis of big data: the ability to efficiently process and analyze large datasets opens new opportunities for extracting valuable knowledge and patterns that may not be obvious to humans. - Continuous learning and improvement: models can be updated and improved continuously as new data arrives, increasing their accuracy and effectiveness over time. - Flexibility of application: the ability to be applied in a wide range of fields, from automating household tasks to complex problems in medical diagnostics and financial analysis.

Disadvantages of machine learning:

Dependence on data: the quality and volume of available data directly affect model performance. Incomplete, distorted, or biased data can lead to inaccurate or biased conclusions. - Difficulty of interpretation: many machine-learning models, especially in deep learning, are “black boxes,” which makes it difficult to understand how a particular conclusion or prediction was made. - High requirements for computational resources: training complex models, especially on large volumes of data, may require substantial computing power and time. - Problems with overfitting: models may “memorize” the training data too well, losing the ability to generalize and to work effectively with new data. - Ethical and social issues: the use of machine learning raises questions of privacy, security, bias, and fairness, especially when models are used in critical decision-making systems.

The machine-learning approach offers significant advantages for developing AI systems that can learn and adapt, but it requires a careful approach to data selection and processing, model interpretation, and consideration of ethical and social consequences.

8.1.3 The Deep-Learning-Based Approach

The deep-learning-based approach is one of the most important and dynamically developing areas in modern artificial intelligence. Deep learning is a subfield of machine learning that uses multilayer neural networks to analyze and process data. The main idea is that each layer of the neural network extracts and transforms the data, gradually producing an increasingly abstract representation. This allows the model to identify complex relationships and patterns that are difficult or impossible to detect using traditional methods.

Artificial neural networks are inspired by the structure of the human brain and consist of individual computational units—neurons. These neurons are organized into layers. The first layer receives the input data; the last layer produces the result; and intermediate, or hidden, layers perform most of the processing. In deep learning, the presence of many hidden layers makes it possible to build more complex and flexible models.

One of the most illustrative examples of deep learning is image recognition. Early approaches to computer vision required manual selection of features such as edges, corners, or textures. Deep neural networks can learn these features automatically. In the first layers, the model may recognize simple lines and contours; in deeper layers, it can recognize parts of objects; and at the final stages, it can identify whole objects, such as faces, cars, or animals.

Deep learning has also had a major impact on natural language processing. Traditional methods often relied on manually designed rules and used complex algorithms to analyze structure and semantics. Deep learning, in turn, can analyze sequences of words while taking context and semantic relationships into account, which makes it possible to create accurate and natural systems for translation, text generation, and analysis.

The main advantage of deep learning over other approaches lies in its ability to automatically find and use the most relevant features in data without the need for preliminary human annotation.

Nevertheless, despite all its advantages, deep learning also has its difficulties. It requires substantial computational resources and large volumes of data for training. However, as technologies develop and computing power becomes more accessible, this approach is becoming increasingly available and widely used in various fields.

Modern achievements in this area have already demonstrated its potential in medicine, robotics, the automotive industry, and many other fields. This makes it possible to solve tasks that were previously considered unattainable for machines. At the same time, deep learning opens new horizons and opportunities for researchers, allowing them to create systems capable of more complex and deeper forms of data analysis. It is possible that deep learning may become the key to creating the artificial intelligence of the future—an AI that is even more powerful, flexible, and capable of independent learning.

8.1.4 Evolutionary Algorithms and Genetic Programming

Evolutionary algorithms are metaheuristic optimization algorithms inspired by biological evolution. They are used to find approximate solutions in complex optimization problems where classical methods may be ineffective.

Evolutionary algorithms are based on three key principles:

Variability is the ability of individuals in a population to change. In the context of evolutionary algorithms, this appears through the processes of mutation and crossover.

Mutation: This is the process of randomly changing the characteristics of an individual. Mutation introduces novelty into the population, which can help the algorithm escape local optima.

Crossover: This is the process of combining the characteristics of two individuals to create offspring. Through crossover, useful characteristics of different individuals can be combined, creating new individuals that may be better adapted.

Heredity ensures the transfer of genetic information from parents to offspring. In the context of evolutionary algorithms, individuals “inherit” a combination of the characteristics of their parents, which ensures preservation of useful characteristics in the population.

Natural selection is the process by which the best-adapted individuals preferentially pass on their genetic information to the next generation. In evolutionary algorithms this is achieved by evaluating the fitness function of each individual. Those individuals that have better fitness values have a greater chance of being selected for reproduction.

Genetic programming, as a variety of evolutionary algorithms, focuses on the automatic creation of programs or functional blocks that solve a specific task. Thus, instead of manually coding a solution, genetic programming makes it possible to “evolve” the most optimal solution by exploring many possible variants.

Examples of use include:

Development of game strategies: In computer games such as Dota 2, genetic algorithms can be used to train neural networks in optimal strategies, allowing artificial intelligence to handle human opponents effectively or adapt to their level. - Robot control: In robotics, genetic programming can help create robot-control algorithms, allowing robots to adapt to new environments or perform complex tasks without prior programming. - Optimization of design solutions: In engineering and design, genetic algorithms can help automatically create optimal designs, whether efficient building layouts, resource allocation, or optimization of production lines.

The main advantage of genetic programming and evolutionary algorithms is their ability to work in situations where traditional mathematical methods may be ineffective or too complex. This makes them an ideal tool for solving many AI-related problems.

8.1.5 Reinforcement Learning and Agent-Based Learning

Reinforcement learning and agent-based learning represent one of the most intriguing and promising directions in AI. Inspired by learning processes in living beings, these methods differ from traditional supervised or unsupervised learning. Instead of simply learning from provided data, agents trained through reinforcement rely on an interactive process in which they make decisions, receive feedback in the form of rewards or penalties, and adjust their strategies on the basis of that feedback.

One of the most vivid examples of the application of reinforcement learning is DeepMind’s AlphaGo. This system was trained to play the ancient and complex game of Go using a combination of deep learning and reinforcement learning. It is notable that AlphaGo did not merely study games played by professionals, but also competed against itself millions of times, learning from its mistakes and successes. As a result of this intensive training process, the system was able to defeat the world champion in Go, something previously considered almost impossible for a machine.

In the automotive industry, these methods are used to train autonomous vehicles. Instead of simply following preprogrammed routes, autonomous vehicles trained using reinforcement can adapt to new traffic conditions, account for the behavior of other road users, and make decisions on the basis of unforeseen situations.

Another field in which reinforcement learning shows impressive results is the creation of autonomous robots. Instead of programming every movement of a robot, researchers allow it to “explore” the surrounding environment, try different strategies, and learn from the results it obtains. This may include training a robot to walk, manipulate objects, or perform complex tasks such as arranging dishes or assembling parts.

However, despite its potential, reinforcement learning also faces a number of challenges. One problem is the need for a large number of iterations for training, which can be resource-intensive. In addition, configuring rewards and penalties so that the agent actually learns what is required of it and reasonably balances exploration of new strategies with exploitation of already learned ones can be a difficult task. Nevertheless, given the successes in this area and ongoing research, reinforcement learning and agent-based learning continue to develop, promising new breakthroughs in the creation of intelligent systems capable of independently making complex decisions under uncertainty.

8.1.6 Hybrid Systems: Combining Different Approaches

Hybrid systems have become popular because of their ability to combine the best aspects of different approaches to creating AI. They are combinations of several methodologies and technologies, allowing developers to benefit from the strengths of each approach while minimizing their limitations.

Consider an example in medical diagnosis. When a physician encounters ambiguous symptoms in a patient, the physician may rely on experience and intuition, as well as on medical literature and databases. A hybrid AI system in this situation may use symbolic programming for logical processing of knowledge from medical textbooks, deep learning for the analysis of medical images, and reinforcement learning for modeling the physician’s decision-making process. Thus, instead of relying on only one method, the system uses a combined approach that provides higher diagnostic accuracy.

Hybrid systems can also be useful in the financial sector. Financial markets are complex systems in which traditional algorithmic methods can work together with machine-learning methods to predict market trends or assess risks. While symbolic programming can be used to define strict rules and policies, deep-learning methods can analyze large volumes of data, such as news articles or social media, in order to detect hidden patterns or trends.

It should be taken into account, however, that creating hybrid systems requires a deep understanding of each approach used. Developers need to carefully weigh the advantages and disadvantages of each method in order to determine how and when they should be combined. Despite the difficulties, hybrid systems offer a promising perspective, enabling the creation of more powerful and adaptive AI-based solutions that can successfully cope with diverse and complex tasks.

8.1.7 Neoclassical Approaches and New Trends in AI

Neoclassical approaches to AI are often based on ideas that move away from traditional methods such as symbolic programming or classical machine learning. They seek to recreate intellectual processes occurring in the human brain through alternative models and methodologies.

One vivid example of a neoclassical approach is the creation of cognitive architectures. These architectures attempt to reproduce the human thinking process by modeling different cognitive functions in an integrated system. For example, the ACT-R project, which began at Carnegie Mellon University, aims to recreate human thinking through a comprehensive cognitive model that can learn and make decisions in various situations.

It is also worth mentioning the new wave of interest in neural-network architectures that do not simply imitate the structure of the human brain, but also try to reproduce its operation at a deeper level. These architectures explore how the human brain processes information, learns, and adapts to new conditions, and they attempt to recreate these processes in machine code.

In addition, recent years have seen a surge of interest in researching and creating emotional intelligence in machines. Understanding and interpreting human emotions is becoming important in the development of new artificial-intelligence systems, especially in the field of human-machine interaction. For example, some studies are aimed at creating robots capable of recognizing and responding to a person’s emotional state, which may be useful in medicine or psychotherapy.

In conclusion, neoclassical approaches and new trends in AI expand the horizons of what is possible and make it possible to create systems that are even closer to human intelligence. These methods and studies leave us with the next question: if we continue moving in this direction, how close can we come to recreating the human mind in a machine?

8.2 Why We Need Knowledge about the Similarities and Differences between Human Intelligence and AI

As AI technologies develop, it becomes increasingly evident that systems possess fundamentally different cognitive qualities and capabilities compared with biological systems, including the human brain. This gives rise to a number of questions and proposals, among which is the need to understand the advantages and disadvantages of each system and how they can be used most effectively.

Understanding the differences between humans and machines becomes the key to creating an optimal interactive system. Artificial intelligence can process enormous volumes of data much faster than a human, predict outcomes on the basis of complex algorithms, and perform routine tasks without fatigue. Meanwhile, human intelligence is distinguished by the capacity for creative thinking, intuition, and social interaction.

Defining the boundaries between a human decision and a decision made by AI becomes necessary in modern society. Which tasks can be entrusted to AI, and in which situations is human intervention required? The answers to these questions may determine the future of many professions and, in a broad sense, even the social structure of society.

Knowledge of similarities and differences will also help us understand whether we should strive to create AI “partners” with human-level intelligence or instead focus on complementing and compensating for human limitations. Thus, the more we learn about the mechanisms of AI, the better we can use it to our benefit.

In addition, for effective collaboration between people and machines, it is important to develop “Intellectual Awareness” among people. This awareness is needed not only for specialists but also for the general public. After all, artificial intelligence is becoming part of our everyday life, and the more we understand its features, the safer and more productive our interaction with it will be.

Thus, knowledge about the similarities and differences between human intelligence and machine intelligence not only helps us better understand our world, but also determines how we will build our future in an era of technological innovation.

8.2.1 Do We Need Human-Like Artificial General Intelligence (AGI)?

With the technological development of society, an inevitable question arises: do we need artificial general intelligence, or AGI? To answer this question, it is important to understand what AGI is and what its potential advantages and disadvantages are.

Artificial general intelligence, or AGI, refers to systems that can autonomously and effectively achieve complex goals across a wide range of environments, similarly to the human mind. However, it should be noted immediately that most modern definitions of AGI are based on an anthropocentric (human beings are the center of the universe) mode of reasoning, which may be unjustified.

First, the concept of “human-like” AGI may be mistaken. Human intelligence is only one of many forms of intelligent behavior in the natural environment. The pressing question is whether we really need AI that copies the human brain exactly, or whether we need intelligence capable of performing tasks effectively regardless of how “human” it is.

Second, the use of AGI with a greater degree of autonomy requires a deep understanding of interaction between humans and machines. A human professional working with AGI must have a clear understanding of how AI “thinks” and “acts” in order to ensure effective cooperation.

On the other hand, AGI with the potential for genuine cooperation based on shared goals and mutual understanding may bring undeniable advantages in various industries. Examples include complex scientific research, big-data analysis, and solving global problems such as climate change. In these contexts, AGI may act faster and more effectively than any human being or team of specialists.

Nevertheless, the risks should be remembered. With the development of AGI comes the possibility that such systems may become so autonomous that they act outside human control. In addition, issues of security, privacy, and ethics become even more relevant as AGI capabilities grow. Should AGI be given the right to make decisions that may concern life and death?

In striving to create systems with human intelligence, we assume that human intelligence is the true form of reason. This idea is already present in the very name “artificial intelligence,” as if it were less real than biological intelligence. We know ourselves as the most intelligent beings in the universe and see ourselves as entities capable of solving complex problems using logic, experience, intuition, and statistical analysis. Yet it is difficult for us to accept the thought that we may be less intelligent than we think. As artificial intelligence develops, we constantly change our understanding of what true intelligence is, limiting it only to human abilities.

In order to understand the difference between human and artificial intelligence, three main ideas should be taken into account:

We may not be as intelligent as we believe when it comes to cognitive tasks. Why, then, are we striving so actively to create artificial intelligence resembling human intelligence? - There are many forms of intelligence, and general intelligence does not necessarily correspond to human-level intelligence. - There is often no need to create general-purpose artificial intelligence. Many complex tasks can be solved with several narrowly specialized artificial intelligences.

Therefore, returning to the original question: do we really need AGI? The answer is ambiguous. On the one hand, AGI may bring unprecedented advantages in innovation and progress. On the other hand, the pursuit of “human-like” AGI may not only be technically difficult but may also raise ethical problems. This means that first we need to understand the goals we are setting for ourselves. Perhaps, instead of creating “human” AGI, we should strive to develop intelligent systems that act in humanity’s best interests without necessarily copying us in every aspect.

8.2.2 The Intelligence of Homo sapiens

What is our real level of intelligence? The answer to this question depends largely on the point of view from which we consider the problem, as well as on the criteria and measures of intelligence that we choose. For example, we can compare the nature and abilities of human intelligence with those of other animals. In such a comparison, we look very intelligent. Thanks to our enormous capacity for learning, we have the broadest range of cognitive abilities for independently solving complex problems and achieving complex goals. Thus, we can solve many arithmetic, conceptual, spatial, economic, social-organizational, political, and other problems. It follows that, compared with other animals known to us, we can indeed consider humans highly intelligent beings.

But we can consider our intelligence not only from a “relative interspecies” point of view, but also try to define it in more absolute terms. For example, we can consider the computational capabilities of the human brain as a physical system. Among AI researchers, the prevailing view is that intelligence ultimately comes down to information and computation. Theoretically, there are no physical laws prohibiting the creation of physical systems with computational power and intelligence exceeding that of the human brain. This means that machines may one day become much smarter than we are. Compared with other animals, our intelligence is high; but in absolute terms it may be very limited, especially because of the limited size of our brain.

To better understand our biological intelligence, we should consider the evolution of our biological thinking abilities. Our brain, as a biological neural network, has undergone an evolutionary optimization process for more than a billion years. It has become a highly efficient system for regulating basic biological functions and for performing perception and pattern-recognition tasks. Throughout our evolution, the brain’s neural networks have been optimized to perform these basic tasks. However, we should not forget the high computational complexity of these processes. For example, when we tie our shoelaces, millions of signals are transmitted through various sensory systems of our body. Such basic abilities developed over millions of years. Our cognitive abilities began to develop much later in evolution. This recent layer of human achievement is built on “ancient” neural mechanisms necessary for survival. Therefore, it is not surprising that our abilities to perform new cognitive functions are still limited. These limitations appear in various ways:

We can consciously process only a limited amount of cognitive information. - Most cognitive tasks, such as reading or calculation, require our full attention. - Although we can process a great deal of information in parallel, we cannot simultaneously perform cognitive tasks that require reflection. - Acquired cognitive knowledge and skills deteriorate over time much faster than perceptual-motor skills.

Our limited ability to process cognitive tasks is not the only factor determining our cognitive intelligence. Human processing of cognitive information displays systematic distortions. These appear as cognitive biases.

Cognitive biases are systematic tendencies or inclinations that distort information-processing processes, making their outcomes inaccurate or erroneous. Many of these biases appear almost identically in different decision-making situations. More than 200 cognitive biases have been described. Most of them are unconscious and seem natural to us, which is why they are often called “intuitive” or “irrational.” However, such distorted thinking can lead to less optimal decisions.

These biases are mainly caused by the properties of the brain as a neural network. For example, lateral inhibition (the ability of an excited neuron to reduce the activity of its neighbors) increases differences in neural activity, which is useful for perceptual-motor functions. But “higher” cognitive functions, such as conceptual thinking or probabilistic reasoning, appeared relatively recently in evolution. These functions are based on ancient neural mechanisms, which can lead to distortions in cognitive information processing.

Since many of these biases are based on neural or evolutionary processes, simple methods such as training or mindfulness courses are unlikely to eliminate them.

8.2.3 The Difference between Artificial Intelligence and Human Intelligence

AI research is often conducted from an anthropocentric point of view, in which human intelligence is regarded as the sole and obvious model. However, the differences between machine intelligence and human intelligence appear in many aspects.

Basic structure: Biological intelligence, based on carbon, rests on neural “living hardware” that is fundamentally different from silicon-based artificial intelligence. Unlike biological “living hardware,” in silicon or digital systems the “hardware” and “software” components function independently of one another. When a biological system masters a new skill, it remains integrated into that system. By contrast, if an AI system masters a certain skill, the corresponding algorithms can be copied to all similar digital systems.

Speed: Signals transmitted by AI systems propagate practically at the speed of light. In humans, the transmission speed of nerve impulses does not exceed 120 m/s, which is orders of magnitude slower than modern computing systems.

Connectivity and communication mechanisms: Homo sapiens does not have the ability for direct information exchange. Human communication processes are implemented through verbal and nonverbal channels with limited bandwidth. In the context of artificial intelligence, such processes are carried out with high efficiency through direct connections, allowing algorithms to be integrated for joint operation.

Updatability and scalability: Algorithmic systems are practically free of limitations in terms of updating, scaling, and reconfiguration. This gives them the ability to maintain up-to-date algorithmic solutions and corresponding data-processing parameters. Such rapid structural transformation and immediate adaptive improvement are the prerogative of artificial systems and are not characteristic of biological objects.

Energy efficiency: At the same time, biological systems have unique energy efficiency. The neural structures of the human brain demonstrate striking energy efficiency, surpassing modern computing devices by several orders of magnitude. As an illustration, the energy consumption of the human brain is comparable to the power of a small light bulb, whereas high-performance supercomputers require enough energy to supply electricity to a small settlement.

These differences in basic structure, speed, connectivity, and other aspects lead to substantial differences in the qualities and limitations of human and artificial intelligence. Awareness and understanding of these differences are important for the successful application and integration of AI into our everyday life.

Educational Publication

V. I. Milushkov N. N. Limansky I. M. Zhukovsky [et al.]

Artificial Intelligence: From Theory to Practice: Textbook / V. I. Milushkov, N. N. Limansky, I. M. Zhukovsky [et al.]; edited by S. O. Baryshnikov

Textbook

Edited by S. O. Baryshnikov

198035, Saint Petersburg, Mezhevoy Canal, 2

Tel.: (812) 748-97-19, 748-97-23

E-mail: izdat@gumrf.ru

Responsible for publication
Editor | S. O. Baryshnikov
Original layout

Signed for printing on 00.00.2024

Format 60×90/16. Offset paper. Typeface: Times New Roman

Standard printed sheets: 0.00. Print run: 00 (first batch — 00) copies. Order No. 000/24