Artificial intelligence. Ch2

Anecdote from a magazine. One man bought a house with artificial intelligence. Within a week, the smart home called him lazy, and a month later the man himself washed the dishes and socks, after which football was turned on for him.

The concepts of natural intelligence and artificial intelligence (AI, AI from the Latin intellectus – knowledge – understanding, reason), the ability of thinking, rational cognition, in humans – AI, in robots – AI. AI can be defined as a branch of computer science that deals with the automation of intelligent behavior in non-living objects. Here we will not evaluate and analyze numerous other definitions of AI and focus on the texts proposed by different authors, so as not to get stuck on this. The understanding of AI as a system capable of solving problems that in the past were only possible for humans, without any mention of consciousness emulation, is also used. And modern AI systems fully meet this definition.

In addition, AI must implement the main property – the ability and desire for self-learning without human participation via the Internet. Artificial intelligence, first of all, is independent learning and independent thinking based on information received and knowledge as a result of self-learning.

Neither approaches nor methods are disclosed here in detail, nor are any details specified in specific AI systems. Each of the elements listed in the article can be requested on the Internet and a detailed response-description will be received for it from different authors. The goal of the publication is to provide the reader with an in-depth look “from above” at the entire AI problem.

The purpose of the publication is primarily educational, cognitive, popularization of science, as well as the desire to attract an influx of new young minds into the ranks of researchers and science, to arouse in such minds a desire to find answers to emerging questions. The scale of the topic requires the introduction of reasonable restrictions on the material presented after a brief panoramic consideration of it.

General provisions about AI

Artificial intelligence is a branch of modern science that studies ways to teach a computer, robotic technology, or analytical system to think intelligently and act just like a human.

Main goals AI should be named and recognized:
– Creation of analytical systems that have intelligent behavior, can learn independently or under the guidance of a person, make predictions and build hypotheses based on an array of data.
– The implementation of human intelligence in a machine is the creation of robotic assistants that can behave like people: speak, think, learn, understand and reason, and perform assigned tasks.

Basic approaches used in AI development:
– descending (English Top-Down AI), semiotic – the creation of expert systems, knowledge bases and logical inference systems that simulate high-level mental processes: thinking, reasoning, speech, emotions, creativity, etc.
– ascending (English: Bottom-Up AI), biological – studying neural networks And evolutionary calculationsmodeling intelligent behavior based on biological elements, as well as the creation of corresponding computing systems, such as neurocomputer or biocomputer. The approaches are united by a common goal and principles.

In addition to those mentioned in publications, you can find others: the most general approach assumes that AI will be able to exhibit behavior that is no different from human behavior, and in ordinary situations (a machine will be intelligent when it can pass the Turing test); science fiction writers are developing their approaches. the basis is the machine's acquisition of intuition, human feelings and creative abilities.

Are being considered AI types:

– Artificial intelligence narrow purpose (Artificial Narrow Intelligence ANI). Artificial Narrow Intelligence (ANI) is also commonly known as Weak AI or Narrow AI. It is an already human-made AI, a goal-oriented tool with a limited perspective that performs specific goal-oriented tasks without the ability to self-reveal the mechanism (functionality). Machines that are focused on one narrow task operate under a narrow set of constraints, which is why they are usually called weak AI.

– Artificial intelligence general purpose (Artificial General Intelligence AGI). For this purpose, initially AGI (often called strong AI) is not trained and does not have any knowledge about the environment, but the device is equipped with a variety of sensors. Power is supplied to the input of a device equipped with AGI, a goal is formulated, and the output receives control actions on the executive bodies of the device, the functioning of which leads to the achievement of the goal. This should be carried out for arbitrary purposes and any environment. AGI is an inverse problem solver.

A clear example of strong AI is the game Detroit: Become Human. In this game, robots are as close as possible to people, they think, feel, learn, are aware of their own “I”, and are able to make decisions almost like people. “Almost”, because neither Alice nor Siri can think independently and make decisions in unprogrammed situations. Strong AI is still an unfulfilled dream.

– Artificial superintelligence (Artificial Super Intelligence ASI is a hypothetical AI that could do everything a human can do. We are talking about computers that surpass humans in their level of intelligence. It is usually mentioned that he must pass the Turing test in its original formulation (it is believed that people pass it?), recognize himself as a separate person and be able to achieve his goals.

The main concepts of AI (in international designations) include: Artificial Intelligence (AI – Artificial Intelligence), Machine Learning (ML – Machine Learning) – section AI, Data Science (DS – Data Science) – these are the three main concepts of AI.

Principles of development, creation and functioning of AI

The basic principle of AI is to combine a large amount of data with the ability to quickly, iteratively processing and intelligent algorithms, which allows programs to automatically learn based on patterns and features contained in the data. The Asilomar Principles (2017) are well known – these are 23 recommendations that are important to adhere to when working with AI in order to use it in a positive way.

At the same time, these principles do not offer any specific measures to prevent or correct adverse situations that arise, which means their practical value is questionable. They are only advisory in nature and very abstract

When developing, creating and implementing AI, they are guided by principles, on the one hand ethical, on the other – technological

First (ethical principles) was formulated in 1942 as the laws of robotics by Isaac Asimov in his novel “Round Dance”:
– A robot or system with artificial intelligence cannot harm a person through its action or, through its inaction, allow harm to come to a person.

– A robot must obey the orders it receives from a person, except those that contradict the First Law.

– A robot must take care of its safety, if this does not contradict the First and Second Laws.
In 1986, Isaac Asimov added one more to the laws of robotics. The writer preferred to call it “zero”.
– A robot cannot harm a person unless it proves that in the end this (harm) will be useful for all humanity.

TO technological principles AI developments include:
– Machine learning (ML) is the principle of AI development based on self-learning algorithms. With this approach, a person is only required to load an array of information into the “memory” of the machine and formulate goals. There are several ML methods: supervised learning – a person sets a specific goal, wants to test a hypothesis or confirm a pattern. Learning without a teacher – the result of intellectual data processing is unknown – the computer independently finds patterns, learns to think like a person.

– A neural network is a mathematical model that simulates the structure and functioning of nerve cells in a living organism. Accordingly, ideally it is a self-learning system. If we transfer the principle to a technological basis, then a neural network is a set of processors that perform one task in a large-scale project. In other words, a supercomputer is a network of many ordinary computers.

– Deep learning is a method used to search and discover patterns in huge amounts of information. For such work, which is impossible for a person, the computer uses advanced techniques.

– Cognitive computing is one of the areas of AI that studies and implements the processes of natural interaction between a person and a computer, like the interaction between people. The goal of artificial intelligence technology is to completely imitate human activity of the highest order – speech, imaginative and analytical thinking.

– Computer vision – this area of ​​AI is used to recognize graphic and video images. Today, machine intelligence can process and analyze graphical data and interpret information in accordance with the surrounding environment.
– Synthesized speech. Computers can already understand, analyze and reproduce human speech. We can already control programs, computers and gadgets using speech commands. For example, Siri or Google assistant, Alice in Yandex and others.

In addition, it is difficult to imagine the existence of artificial intelligence without powerful graphics processors, which are the heart of interactive data processing. To integrate AI into various programs and devices, API technology is required – application programming interfaces. Using the API, you can easily add artificial intelligence technologies to any computer systems: home security, smart home, CNC equipment, etc.

AI Timeline

1832 S. Korsakov invented punched cards and 5 intelligent machines
1835 Ch. Babbage chess machine
1914 L. Quevedo device for playing chess
1924 K. Chapek introduced the concept and term “robot” – the play “Universal Robots” London.
1930 Training artificial intelligence as a child
1940 Modeling of thinking, neurocybernetic and logical approaches
1942 Azimov A. formulates the three laws of robotics.
1943 McCulloch W., Pitts W. introduced the concept of a neural network
1948 Wiener N. Book about Cybernetics
1949 Hebb D. The first learning algorithm
1950 Turing A. Publishes an analysis of the game of chess. Turing test. 1 conf. in AI at Dartmouth, USA. Question-answering systems.1 learning algorithm
1954 Software for playing chess
1956 The term “artificial intelligence” is coined
1958 Rosenblatt F. Single-layer perceptron solves the classification problem MARK-1
1960 Research at Moscow State University and the USSR Academy of Sciences.
1960 D. Weizenbaum developed the first virtual interlocutor (computer program) that supports a dialogue with a person.
1960 Widrow B. Solver of prediction and adaptive control problems using memistors
1963 Petrov A.P. Results of solving difficult problems for a perceptron
1964 Maslova S. “The reverse method of establishing deducibility in calculus. predicates.”
1965 Created Eliza, the first robot assistant capable of speaking English.
1966 V. F. Turchin developed the language of recursive functions Refal.
1969 Minsky M. Proof of the limited capabilities of the perceptron
1969 Stanford created Sheki, a moving robot with AI
1970 The backpropagation algorithm. Cognitron. Androids. Robot programming language
1972 Kohonen T., Anderson J. A new type of neural networks (NN) with memory
1973 Khakimov B.V. Nonlinear model of neural network with synapses based on splines.
1973 Freddy the robot with AI and artificial vision is created at the University of Edinburgh
1974 The Kaissa program is the 1st world chess champion among computer programs.
1974 Verbos P., Galushkin A.I. backpropagation algorithm, training
1975 Fukushima introduces a self-organizing network – the cognitron.
1980 Barr and Feigenbaum proposed a definition of artificial intelligence (AI).
1982 Hopfield J. Energy minimizing feedback loop.
1986 Development of the backpropagation method.
1989 The chess program Deep Thought beats grandmaster Bent Larsen.
1990 Convolutional neural networks Unsupervised learning. Machine translation, face recognition.
1997 IBM Deep Blue beats world chess champion Garry Kasparov
2000 Torch Library. NVidia Cuda. Expert systems based on neural networks. Deep Learning.T.
2007 Hinton J. Deep learning algorithm.
2008 The beginning of technological singularity, integration of man with computers, biotechnology.
2010 Virtual assistants. DeepFace facial recognition systems. AI for strategy games.
2013 Research on Image Sorting.
2016 Claim. Google Deep Mind has defeated a professional Go player for the first time.
2017 Gamalon introduced self-learning technology based on data fragments.
2017 The European Charter for Robotics was adopted by the European Parliament
2018 Created univers. syst. understanding of NL GLUE – General Language Understanding Evaluation.
2019 The national strategy for the development of AI in the Russian Federation was approved (RUB 90 billion).
2020 National project “Artificial Intelligence” of the Russian Federation.
2023 Update of the National AI Development Strategy.
2024 It is planned to allocate about 5.2 billion rubles for developments in the field of AI in the Russian Federation.

Approaches to creating AI

Top-down approach to creating AI

Proponents of this approach (Minsky, Papert) followed the path of creating programs based on general-purpose computers aimed at (creating symbolic systems) solving intellectual problems: thinking, reasoning, speech, emotions, creativity
– proof of theorems;
– pattern recognition;
– games (chess, etc.); etc.

The methods underlying such programs may be completely different from those that are actually used by humans. They are based on certain forms of knowledge representation, borrowed from experts in specific subject areas and inference mechanisms that implement the problem solving process.

In general, the following approaches are distinguished, which are classified as top-down:
– intuitive;
– symbolic;
– logical

– knowledge is formalized and mechanisms for interpreting this knowledge are built (the algorithm is born in the process of work);
– declarative type systems (knowledge is concentrated not in the program, not in the procedure, but in information structures; the program is only a universal solver)

How is data different from knowledge? Knowledge comes from information, and information comes from data. For information to become knowledge, it is necessary to perform, at a minimum, the following actions:
– Comparison with other elements.
– Prediction of consequences.
– Search for connections, intersection points.
– Communications with other knowledge carriers.

Forms of knowledge representation:

– production rules (if [ ]That [ ]);
– hierarchical structures (frames);
– semantic networks;
– logic algebra apparatus for describing and interpreting knowledge (first-order predicates);

Mechanisms for interpreting knowledge;
Structure of a knowledge-based system

Bottom-up approach to creating AI

The ideas of the bottom-up approach turned out to be consonant with the ideas of neurophysiologist V. McCulloch about how the human brain works (the study of neural networks). In 1943, McCulloch, together with mathematician W. Pitts, developed a theory of brain functioning. In it, they proceeded from the neural organization of the nervous system of a living organism and the binary law of neuron behavior (1 – active, 0 – passive).

N. Wiener’s idea about the principle of feedback is used: “any artificial system that claims to be rational (intellectual), like all living things, must have the ability to achieve certain goals and adapt, i.e. learn.”

• The functions of higher nervous activity are modeled using methods of mathematical logic.
• The possibility of constructing logical networks (automata) that simulate neutron networks (natural), characterized by certain physiological properties, is shown.
• However, the accepted purely logical scheme of interaction between neurons did not correspond to the true processes occurring in the nervous system of a living organism.

So the bottom-up approach:
– It comes not from formalized knowledge, but from the mechanisms of human thinking;
– Modeling neural networks;
– Architecture and algorithms for training neural networks;
– Application of these algorithms in practice (neural computers).

The following approaches are distinguished, which are classified as top-down:
– structural;
– quasi-biological;
– evolutionary;

AI as a scientific field is closely related to such sciences as psychology, biology, medicine, and linguistics. Specialists in these fields build and programmatically implement, together with mathematicians, more and more new models. Based on them, researchers in the field of AI are trying to restore specific forms of manifestation of intelligence. The results obtained, in turn, provide new food for thought for specialists in these fields.

The fundamental difference between the two approaches to building artificial intelligence (AI) systems is what exactly is taken as the basis when creating AI.
The top-down approach is to try to simulate higher psychological functions and thereby achieve artificial intelligence. The bottom-up approach attempts to achieve the same by modeling the underlying structures and increasing their complexity [Nilsson 1998].

The logical approach to inventing AI systems is based on modeling reasoning and symbolic representation of information. The theoretical basis of this approach is logic, the rules of inference. Another hybrid approach implies that only a synergistic combination of neural and symbolic models can achieve the full range of cognitive and computational capabilities. For example, expert inference rules are generated by neural networks, and generative rules are obtained using statistical learning.

In the hybrid paradigm there are
– Agent-based approach, which is based on the use of intelligent (rational) agents. Here intelligence is the computational part (roughly speaking, planning) of the ability to achieve the goals set for an intelligent machine. This machine becomes an intelligent agent that perceives the world around it using sensors, and the ability to act on objects in the environment through actuators.

The listed paradigms and approaches are implemented through a variety of methods:

• search in state space;

• natural language processing;

• knowledge representation;

• expert systems and decision support systems;

• machine learning and artificial neural networks;

• genetic algorithms;

• multi-agent systems.

Conclusion

Today, the interdisciplinary direction “Artificial Intelligence” is quite developed and methodologically rich. There are a large number of approaches and methods for building AI systems.
Recognition of the capabilities of artificial intelligence can somehow reduce human dignity. Disagreements on the issue of creating artificial intelligence have a pronounced emotional overtones (background). But should the problem of artificial intelligence abilities be combined with the question of the development and improvement of the human mind?

Literature

1. Mirkes E. M., Neurocomputer. Draft standard.]- Novosibirsk: Science, Siberian Publishing Company RAS, 1999 .- 337 p. ISBN 5-02-031409-9 (Chapter 9: “Contraster”
2. McCulloch W.S., Pitts W.,Logical Calculus of Ideas Relating to Nervous Activity // In: “Automata”, ed. K.E. Shannon and J. McCarthy. – M.: Foreign publishing house. lit., 1956. – pp. 363–384. (Translation of the 1943 English article)
3. W. S. Anglin and J. Lambek, The Heritage of ThalesSpringer, 1995, ISBN 038794544X online
4. Bacon, Francis The Advancement of LearningBook 6, Chapter 1, 1605. Online here.
5. John Searle. Is the mind of the brain a computer program?
6. Roger Penrose. The new mind of the king. About computers, thinking and the laws of physics. Publisher: URSS, 2005 ISBN 5-354-00993-6 6. http://www.thetech.org/robotics/ethics/index.html
7. http://region.computerra.ru/offline/2000/42/5284/
8. Harry Harrison. Turing choice. Publisher: Eksmo-Press, 1999, 480 pp.
9. The future of artificial intelligence / Ed. K. Levitin, P. Khoroshevsky. – M.: Nauka, 2003.
10. Zaitsev I.M., Fedyaev O.I. Agent-based approach to modeling intelligent distributed systems: Coll. – Donetsk: DonNTU, 2008. – P. 337-338. 11. Flach P. Machine learning. – M.: DMK Press, 2015. – 400 p. — ISBN 978-5-97060-273-7.