Neural networks, genetic algorithms, etc. ... Myths and reality. Version II
First of all, I would like to thank everyone who gave their critical comments on the first version of the article. It seemed to me that it would be a good idea to write version two, and not just to leave everything as it is.
Of course, artificial intelligence already exists! If you look at the headlines of articles in popular media, the titles and slogans of various scientific conferences on this topic - certainly it is. It is impossible not to believe, especially when you really want to finally find yourself in the 21st century - the “present”, as described in all science fiction novels. But is it? And if not, then what really exists. In an attempt to understand the myths and realities, this article was written.
Initially, I wanted to start something like this: “for the first time the mention of the term“ Artificial Intelligence ”appeared in D. McCarthy in 1956 at a conference at Dartmund University, the founders of AI should be considered W. McCulloch, W. Pits, F. Rosenblat” and t .d However, it is already too late and does not quite meet the objectives of the article, and Wikipedia is ahead of it with such a beginning.
')
Analyzing the latest "victories" of AI, as well as critical articles, one inevitably comes to the conclusion that everything revolves around several common features. One part of the articles criticizes the impossibility of passing the tests, and the other is full of pathos of "incredible victories." At the same time, the fact that victories were achieved in highly specialized tasks is ignored, where the main advantage of the machine was the speed of searching the base of facts and the “ability” to see patterns where people simply get tired of doing it quickly. Brilliant examples of cluster analysis in one form or another and the formation of a base of template facts. All these are consequences, but the causes in most cases are either not analyzed at all, or are considered superficially.
There is no doubt that today AI can be considered a separate science or, at least, a separate branch of knowledge. Any science must have its own axioms, definitions, theorems and hypotheses. And this is where the fun begins. Namely, from the definition of what kind of activity can be called “intellectual” and what can be described as “artificial intelligence”? Next, it is necessary to designate a list of causes and conditions on the basis of which it is possible or impossible to state the fact of the existence of artificial intelligence in its full form, for example, according to Searle. Uncertainty in these issues and generates the entire huge range of myths that have acquired the AI. In addition, there are a number of quite good articles on the history of AI, for example, here or here . But still, a little immersed in the story once again would be appropriate.
Let's start with the definition of what is AI, let's try to follow the drift of the concept itself. Some dates are approximate.
1956 D. McCarthy suggested that “intelligence within this science means only the computational component of the ability to achieve goals in the world,” and instead of the concept software, computational methods.
1976 Newell-Simon’s hypothesis that the ability to perform symbolic computations is quite sufficient to become able to perform meaningful actions. Therefore, all animals and man in fact perform symbolic calculations. To a certain extent, this is a continuation of the mechanism of Descartes and his Cartesian theory. All modern research is built on this hypothesis, in spite of all its criticism.
1971-79 year. Edward Feigenbaum creates the concept of expert systems, which became the prototype of all expert systems. In fact, there was a turning point, after which the current trend of using statistics as the basis for learning was finally fixed as the main one. Research in the field of heuristics was finally pushed to the sidelines.
1980 year. The first fork. John Searle proposed such concepts as strong (full) and weak AI. “... such a program is not just a model of the mind; in the literal sense of the word, it is itself the mind, in the same sense in which the human mind is the mind. ”
1981 Barr and Feigenbaum determined that artificial intelligence is a field of computer science that develops intelligent computer systems, i.e. systems that have the capabilities that we traditionally associate with the human mind - language understanding, learning, ability to reason, solve problems, etc.
1989-1995 The emergence of the term "intelligent agent". An example is the work of Russell and Norvig [1].
2004 Osipov GS [2] belongs to one of the most accurate, from the point of view of the author, definitions of the term and tasks of AI. He defined AI as a set of methods, models, and software tools that allow artificial devices to implement targeted behavior and reasonable reasoning.
Summarizing all the above, I would like to recall Bertrand Russell's remark: everything that seems definite and precise becomes more uncertain with a closer look. The modern understanding of the term AI in fact increasingly goes to interpreting AI as software or Intellectual agents (AI). The main purpose is to solve a certain class of problems in close intertwining with Big Data as a source of facts for analysis. An example is the translation task. For all its apparent comprehensiveness, it is limited to comparison tasks, but it does not solve the problem of answering the questions: “which mountain is the highest?” Or “what if ...”!
Thus, the problem is clearly observed to drift into the area of ​​highly specialized agent systems and about the actual departure from the task of building an AI-complete model towards specialized intelligent agents. This is not bad news, as it removes the likelihood of a catastrophe described by Nick Bostrom, and the very concept of “intellectual” becomes a mere decoration.
As for the main stages, in due time the chronology which came into our eyes describes well the main stages of the formation of AI, unfortunately the author is unknown:
Stage 1 (50's) (Neuron and Neural Networks)
It is associated with the emergence of the first machines of sequential action, with very small, by today's standards, resource capabilities for memory, speed and classes of tasks to be solved. These were tasks of a purely computational nature, for which solution schemes were known and which can be described in some formal language. The tasks for adaptation also belong to this class.
Stage 2 (60s) (Heuristic Search)
In the "intelligence" of the machine, mechanisms for searching and sorting were added, the simplest operations to summarize information that do not depend on the meaning of the data being processed (in this place it is difficult to fully agree with the author). This became a new starting point in the development and understanding of the tasks of automating human activity.
Stage 3 (70s) (Knowledge submission)
Scientists realized the importance of knowledge (in scope and content) for the synthesis of interesting algorithms for solving problems. This meant knowledge that mathematics could not work with, i.e. Experienced knowledge that is not strictly formal and is usually described in a declarative form. This is knowledge of specialists in various fields of activity, doctors, chemists, researchers, etc. Such knowledge has been called expert knowledge, and accordingly, systems based on expert knowledge have become known as consulting systems or expert systems.
Stage 4 (80s) (Learning Machines)
The fourth stage of development of AI became breakthrough. With the advent of expert systems in the world, a fundamentally new stage of development of intellectual technologies began - the era of intelligent systems - consultants who offered solutions, substantiated them, were capable of learning and development, communicated with a person in his usual, albeit limited, natural language .
Stage 5 (90s) (Automated Machining Centers)
The increasing complexity of communication systems and solved tasks required a qualitatively new level of “intelligence” of supporting software systems, such systems as protection against unauthorized access, information security of resources, protection against attacks, semantic analysis and information search in networks, etc. And the new paradigm of creating promising protection systems of all types has become intelligent systems. They allow you to create flexible environments within which all necessary tasks are solved.
Stage 6 (2000s) (Robotics)
The scope of the robots is quite wide and extends from autonomous lawn mowers and vacuum cleaners to modern models of military and space technology. Models are equipped with a navigation system and all sorts of peripheral sensors.
Stage 7 (year 2008) (Singularity)
Creation of artificial intelligence and self-replicating machines, human integration with computers, or a significant abrupt increase in the capabilities of the human brain due to biotechnology.
According to some forecasts, the technological singularity may already occur around 2030. Proponents of the theory of technological singularity believe that if a fundamentally different from human mind (posthuman) arises, the further fate of civilization cannot be predicted based on human (social) behavior.
Let us return to the problems and myths that have arisen largely due to the large number of writers, as well as in the attempt of different teams of developers to get an investment in their research. In order to expose the myths, not less articles have been written and published. An example is such an article as this or this , but one of the best in terms of figurative presentation of articles, etc. ... To write another article, at first glance, is a rather controversial exercise, if you do not pay attention to some common features. A common feature of all these works is the statement of the fact that all of them are built around criticism of the capabilities of neural networks, as well as the fact that we do not know how the brain works, and that building AIs is not necessarily around the idea of ​​neural models. All this is true, but almost none of the authors are trying to formulate what is needed to solve the AI ​​problem. In the framework of this article, I would like to, albeit conditionally, but try to identify the tasks, without solving which it is impossible to move on.
Before proceeding, I would like to make a few comments.
First, eliminate conversations about clandestine research centers in which some forces hide the true state of affairs. Creating a truly intelligent agent requires solving a number of philosophical problems and developing on the basis of these solutions a new mathematical apparatus, which is simply impossible to hide in full.
Secondly, all further reasoning is by no means a certain statement about the “divinity” or “unknowability” of human nature. The task of this article is to show white spots and show the fact of excessively optimistic statements about the fact of the existence of AI in its full form to date.
What are existing technologies? The total pool of technologies, united under the general term "artificial intelligence systems" is not so great. To him with varying degrees of reliability can be attributed:
It is possible that it will seem to someone that completely different technologies are dumped here "in one pile". But this impression is deceptive, because despite the obvious differences, they all have common features:
The very same learning is characterized by the following main features:
Neural networks
According to Wikipedia, the link to it will be forgiven, neural networks "are a system of simple processors (artificial neurons) connected and interacting with each other". Handsomely. There are various options for implementation such as Bayesian networks, recurrent networks, etc ... The main model of work: the base of images is the transfer function-optimizer.
The most widely used today are limited Boltzmann machines in a multi-layered version. Layering, i.e. Depth is needed to overcome the “XOR” problem. In addition, as shown by Hinton, an increase in the number of layers of hidden neurons makes it possible to increase the accuracy due to the presence of “intermediate” images with minimal difference in each layer. In this case, the closer the intermediate layer to the exit, the higher the specification of the image.
The main purpose of creating neural networks and the resulting task of learning is to eliminate the need for intermediate computational conclusions when analyzing the profile-matrix of incoming signals. This goal is achieved by creating a base of reference profiles, each of which must have a single neuron at its output - a cell of the resulting matrix. Each such neuron is assigned a certain interpretation-result.
The main problem from the point of view of the problem being solved, as well as the actual training, is the noise of incoming matrices of the incoming neurons entering the analysis matrices. Therefore, one of the main conditions is the availability of high-quality training sample. If the training sample is of poor quality, then a high noise level will lead to a large number of errors. However, the large size of the training set can lead to the same result.
To some extent, the work of the neural network can be compared with the work of unconditioned reflexes of living beings [3] with all the attendant drawbacks.
This approach works well for tasks where there is no strong noise, a clear a priori base of reference images. The main task is to choose the most appropriate image from an existing knowledge base. The tasks of forecasting, in this case, will be solved only by extrapolating the existing history without the possibility of creating new entities, i.e. induction with insufficient deduction.
Some may argue that this is not the case. But before rejecting the above, it is necessary to determine what is meant by the term “new entity”? Is this another instance within the framework of the existing vector-class space of the selected subject area or the creation of new area-spaces? One of the following articles will be devoted to the topic of comparing something with something that can be compared and what cannot.
The basic principle of human learning is induction and abduction. Induction is difficult at least because of the initial formulation of the problem - it will get rid of intermediate calculations. A possible objection to the fact that the processes associated with induction, can occur at the learning stage - are weak, since learning is based on several fundamental principles:
Genetic algorithms
According to the same Wikipedia: “The genetic algorithm is a heuristic search algorithm used to solve optimization and modeling problems by randomly selecting, combining, and varying the desired parameters using mechanisms similar to natural selection in nature.”
There are a lot of works by such authors as Panchenko T.V. [4]., Gladkova L.A., Kureichik V.V. [4], etc ... The basics of the “genetic approach” are well disclosed here, many interesting works on the application of genetic algorithms . For example, the work of I.Yu. Pleshakova, S.I. Chuprina [5], V.K. Ivanov’s article. and Mankina PI [6], articles on Habré and a number of others.
One of the most important advantages of genetic algorithms is the absence of the need for information about the behavior of a function and the insignificant influence of possible discontinuities on optimization processes. As in the case of neural networks, there is a departure from the need to analyze cause-effect relationships, by building a "final" image - the objective function. In this sense, from the point of view of solving text analysis, searching, genetic tasks solve the same problems or are very similar to the methods of latent semantic analysis.At the same time, we must pay tribute, in terms of semantic search and text indexing, genetic algorithms have much greater prospects, compared with the methods of latent-semantic analysis.
From the point of view of pattern recognition, with a very strong stretch, the objective function can be compared with the layer of input neurons and with the expectation of a maximum as an analogue of maximizing the signal of the neuron of the output layer. Although it would be more correct to say that genetic algorithms are used to improve the efficiency of learning of neural networks, they still cannot be considered as a competition for neural networks. Tasks are different.
A common drawback - the absence of induction algorithms - is fully present.
JSM methods
The JSM-method of automatic hypothesis generation, proposed by V.K.Finn (Finn, 1983), is a synthesis of cognitive procedures of induction, analogy and abduction [Finn, 1999]. It can be formulated as a system of rules and even presented as a program in a logic programming language [Vinogradov, 1999], [Vinogradov, 2001], [Efimova et al, 2006]. [7]
The saturation condition will be satisfied if the application of the induction procedure does not lead to the generation of new hypotheses. The test for causal completeness is precisely the use of some variant of abduction. The data set and the variant of the strategy of the JSM method satisfy the condition of causal completeness, if all data on the presence or absence of properties of objects are explained using hypotheses about the possible causes of the presence or absence of properties generated by the induction procedure.
If the condition of causal completeness is not fulfilled, then this is the basis for replenishing the base of facts and / or changing the strategy of the JSM method and / or the way of presenting data. All these operations are performed with the participation of the person. From the point of view of V.K. Finn [8] the most promising in this matter is the further development of JSM methods.
One could add a couple of technologies to this list, but this would not affect the picture as a whole.
Thus, today the term “AI” rather means a subspecies of technological (algorithmic) approaches to solving combinatorial problems. The main tasks of which are reliable separation of “statically significant” in its essence patterns and building images-objects on the basis of statistics, without analyzing cause-effect relationships. The main directions are pattern recognition. Under the images can be understood images, sounds, a combination of symptoms of disease, etc.
The result of learning neural network, etc. there must be some identified pattern, presented in the form of a certain matrix-cluster (vector). Of course, this matrix or set can be constantly corrected by new examples, but this does not affect the essence of what is happening. In any case, the set revealed and cleared of noise can be represented in the form of “alienable logic”, which is a kind of “optimal” way to solve the problem. An example of such an area is the task of automatically categorizing texts, but not from the point of view of posting texts by already known headings, but the creation of headings itself. Their annotation, as well as the automatic construction of various kinds of ontologies.
According to most experts, all these methods are nothing more than reincarnations of the tasks of statistics and clustering. Therefore, all modern efforts in the field of deep machine learning, etc. no more than the efforts of entomologists to study butterflies. Search and description of the color and shape of the wing does not give the slightest understanding of the nature of flight. At the same time, it should be recognized that they are very useful from the point of view of recognizing images of both visual and other similar objects, whether they are medical histories or analysis of chess games ...
All this leads to the conclusion that further reasoning in the direction of research of the physical or "neural" model of the construction of AI is a dead end. All further reasoning must be built on the basis of logical abstractions such as determining what kind of intelligence is, what signs, what “mechanics” of the processes of deduction, induction, abduction. What is the logic and mathematics of the basic laws of logic. What kind of logic and mathematics is necessary, let's say, to “invent” to create AI. What is necessary to solve the AI-complete problem? To begin with, it is necessary to determine the term "intellectual", "intelligence" and what features it should have.
What is considered to be intelligence and even more artificial? And here the first difficulties begin to arise. For example, to consider whether intellectual systems created for solving specific problems, having the potential for logical conclusions and self-learning within the framework of the task field. If, however, we consider an AI as a system capable of self-awareness and goal-setting, can the term “artificial” be considered correct? Indeed, in this case we will deal with a completely independent person, albeit with a different physical nature. Yet “artificial” is not just man-made, but rather limited to artificial borders. Otherwise, in some way all people are man-made, as they are created by other people.
In accordance with the definition, for example, in Ozhegov's dictionary: “INTELLIGENCE, -a, m. Mind (in 1 value), mental ability, mental beginning in a person. Tall and. II adj. intellectual, th, th. Intellectual abilities Intellectual property (protected by the law of someone's work.). " By definition, D. Wechsler "intelligence - is the global ability to act rationally, think rationally and cope well with life circumstances." In Kant (German: Verstand — reason) as the ability to form concepts, and reason (German Vernunft) as the ability to form metaphysical ideas. This usage has become widespread in the subsequent German philosophy and finally settled in Hegel in his concept of reason (I.) and reason. In accordance with the new philosophical encyclopedia "INTELLIGENCE (lat.intellectus - mind, reason, mind) - in a general sense, the ability to think; in epistemology - the ability to mediate, abstract cognition, which includes such functions as comparison, abstraction, the formation of concepts, judgment, inference; opposes the direct types of knowledge - sensual and intuitive; in psychology - rational, subject to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].subordinate to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].subordinate to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].
Unfortunately, all philosophical definitions, due to their insufficient “mechanism”, do not provide a sufficient basis for solving the problems of constructing AI.
From the point of view of understanding the tasks of AI, the definitions from biology and psychology are closer. For the most part, they boil down to the fact that Intellect is a type of adaptive behavior aimed at achieving a goal. Achievement of goals and adaptation of all researchers is put at the forefront. The analysis of targeting mechanisms is in itself a topic of a separate article, so for now we will limit ourselves to the fact of its presence.
From the point of view of algorithmic problems and logic building, of great interest are the works of Joey Paul Gipford, who in the 1950s proposed a cubic model of the structure of intelligence. Remarkably, world-wide fame brought him research in which he, based on the use of tests and factor analysis, attempted to construct a mathematical model of a creative personality.
The structure of the intellect according to J. P. Gipford.
Separately, A. Turing with his work "Can a Machine Think" and a child machine should be singled out.
The properties of intelligence encountered in different works may already be more specific and understandable from the point of view of algorithms and models for constructing AI. So the list of essential qualities of human intellect, can be quite interesting. These include:
Such properties as “depth”, “proof” and “criticality” are understandable and more or less algorithmized. With the question of the formalization and algorithmization of such properties as “curiosity”, “flexibility and mobility” everything is not so obvious.
To start reasoning about the problems of mathematical logic when creating AI should follow from a formalized presentation of such concepts as “value”, “meaning”, “utility”, “ethical”, etc. ... In fact, there are still no solutions to the problems of formalizing meaning, induction and others complex logical and philosophical questions. For example, what is knowledge, how is it arranged and how is it related to facts. How to calculate the conclusions based on knowledge? What is a sign and what is the minimum unit of meaning? How to operate with predicates of more than first order? Unfortunately, the modern understanding of these concepts inevitably reduces to any functions, clustering, maximizing or minimizing any target parameters. At the same time, differentiability of space is desirable. In reality, thinking processes are not differentiated,rather, they represent something similar to fractals with inno- and self-references. In this sense, for example, the works of Umberto Maturana with his autopoiesis [10], V. Tarasenko [11] or the work of V.F. Turchin, who in 1966 developed the language of recursive functions Refal. Refal was intended to describe languages ​​and different types of processing.
As accurately noted by Pantnam [12]: "concepts ... are not images or mental representations ... but rather abilities and skills that create opportunities for operating with signs for the unity of our behavior and sign systems." And in the language of mathematics, remarkably expressed by Mandelbrot, thinking has an analogy to a fractal. The shape of the fractal does not depend on the sample or the pattern reproduced in the iterations. This form depends on the nature of the infinite iterations of the transformation method [13].
Leaving aside the high matter, we will try to understand at least with such "basic" concepts as "induction", "abduction", "deduction". There are a lot of definitions of all these terms, so we’ll give some comments and try to comment on them. Unfortunately, within the framework of the article it is impossible to consider these issues in detail, otherwise it would have been a book. Therefore, it is necessary to be limited to a superficial overview of some problems. The analysis of such concepts as "knowledge", "fact", "sign", especially from the point of view of algorithms, requires a separate article.
Deduction- inference from the knowledge of a greater degree of community to a new knowledge of a smaller degree of community [14]. From the point of view of algorithmization, perhaps, the most understandable and fully realized concept. All modern rhenium in the field of AI, such as neural networks, just implement this logic.
Induction- inference from knowledge of a lesser degree of community to a new knowledge of a greater degree of community. In mathematical logic, this is a conclusion that gives a probabilistic judgment [15]. The basic problems of induction, formulated by Hume back in 1888 and to some extent solved by Popper K. [16], still / only / already in 1933, were, for example, how to substantiate the belief that the future will be (to a large extent) same as the past? How can one evaluate faith in the fairness of expectations in the future? Is our reasoning justified the transition from cases re-encountered in our experience to other cases with which we have not met before? How to mathematically take into account the habit, as an integral part of the induction process? It would seem that for this there are such things as Bayesian probabilities, etc ... But if you think about it, they absolutely do not give an answer to these questions.It is impossible to make repeatability the main factor in the inductive conclusion and how is the irrationality of behavior? This is only part of the problem. Consequently, a mathematical model based on “classical” statistics does not fully work here ...
The law of absolute quality , formulated by Hegel and populistly paraphrased by Engels as the law of the transition of quantity into quality. In many ways, this law is a refinement of the previous topic. The meaning of the law is that the number of observations is in no way connected with the emergence of new properties. In this case, Engels went to the actual falsification for the sake of his ideas. As well as the problem of induction today has no mathematical solution.
Abduction- cognitive hypothesis acceptance procedure. For the first time clearly highlighted Ch.S. Pierce, who considered abduction (abductive inference) along with induction and deduction. H.S. Pierce believed that by selecting the most essential hypotheses among a vast number of hypotheses, the researchers implement the “abduction instinct”, without which the development of science would be impossible [17]. The problem is that today there is no mathematical apparatus to calculate the threshold, and the rules for accepting hypotheses themselves. How to understand that a certain group of hypotheses explaining groups of facts are “plausible”? Pierce considered the ability of the human brain to such operations "abductive instinct" of man. What is the mathematical model of abduction thinking?
One of the main reasons for the failure of the formalization of thinking processes was formulated by Kant [18] as the problem of the “completion” of artificial systems, in contrast to the natural, which are amorphous in structure and logical “incompleteness”. Thus, all these concepts are quite complex and still not clearly formulated not only from the standpoint of mathematical, but also philosophy as such. Without the implementation of these problems it is impossible to create Intelligent agents capable of solving common problems. But today the author is not aware of this kind of logic. Nick Bostrom’s statement [19] that "... progress on this path requires technological solutions rather than technological breakthroughs ..." seems somewhat more optimistic.
Nevertheless, there are several remarkable studies, such as the works of Finn V.K. in his work “On Data Mining”, [20] and also in [8]. A complete analysis of Finn's ideas would require more than one article, so we restrict ourselves to a brief squeeze.
One of the most important hypotheses, in the author’s opinion, is the logical abstraction proposed by him, which could be called a structural model of any automated intellectual system:
AIIS = (RIS + PIS + ININ) + PAP ,
Where:
AIIS - automated intelligent information system,
RIS = (generator of hepatitis + theorem proving + calculator), Finn V.. called it “Discloser”,
IPR - search and information system,
INIP - intelligent interface (dialogue, graphics, user training in the system)
PAH - subsystem of automatic database replenishment from the “information environment”
AIIS, in turn, has two blocks: a fact database (DB) and a knowledge base (KB). A database consists of objects with a given structure: finite sets, tuples, words, graphs, etc ... The knowledge base consists of: axioms defining the data structure (for example, axioms of Boolean algebra, algebra of tuples ...), connection axioms of the original (for DB) predicates that implicitly define the class of problems to be solved (for example, problems of recognition of causal dependencies).
For the full implementation of full AI according to Surl, there is a lack of understanding where such subsystems as consciousness should be located and the functions that follow from this to determine the purpose of the system or what we call “curiosity”, where to lay the “depth of mind”. Without this, an independent statement of the problem by the system is impossible. We must pay tribute to Finn, in his works he draws attention to this, but within the framework of this abstraction did not consider it necessary to allocate these issues into a separate functional unit.
Analyzing the components, the question arises of the need to solve a huge range of mathematical and analytical problems. With the generation of hypotheses or deduction everything is not so bad, at least there are quite working hypotheses. But as for the philosophical and mathematical questions of induction, abduction is still a complete obscurity. How to work with higher order predicates? What is the general structure of the axioms? How to solve the problems of operating predicates of more than two three orders of magnitude? Create a programming language or operate with something similar to graphs and multidimensional arrays? There is no comprehensible mathematics for work with associative fields. How to analyze the competition of conclusions within the framework of the multidimensionality of space as associative links,so multidimensionality of the space of possible conclusions? Indeed, in each of the “possible worlds” its conclusion will be correct ... What is the logic and mathematics of the problems of inducing quantity and quality? This is only an insignificant part of the issues that are still to be resolved.
As with any scientific hypothesis, the scheme proposed by Finn is not definitive. But today it is perhaps one of the most working hypotheses. The analysis of each of the elements of the scheme itself is worth a separate article.
Studies in the implementation of each of the components, creating a highly efficient bus for exchanging data between these components can be quite promising and exciting tasks. Under the conditions of rapid growth of SaaS, cloud computing, the creation of such systems, built by integrating components from different teams, could give a significant breakthrough in this matter. Intellectual agents created on this principle could really learn to solve a much larger class of problems.
The last remark that I would like to make concerns the term “Artificial Intelligence”. It would be fair to understand the term AI as a general and, rather, a philosophical concept. Within the framework of practical tasks, the use of the term “Intellectual agent” would be more appropriate. This term could provide a clearer understanding of the limitations of the solution in terms of the amount of information, the logic used, the spectra of the tasks being solved and the goal-setting of the systems being created. The use of the word “limitedness” should be understood not as impairment, but as an understanding of the limits of possibilities and to avoid possible substitution of concepts.
Solving the tasks themselves ML, the creation of tensor processors is at best nothing more than digging into one of the components of the Finn model.
findings
Currently, none of the problems of logical inference (deduction-induction-abduction) is solved except for a partial solution of issues related to deductive conclusions. From the point of view of complete AI, not a single problem has been solved at all. The main problem is that the overwhelming majority of studies do not affect the implementation of the abstraction itself, at least in the understanding of Finn, and therefore are deadlocked in terms of AI-full implementation.
As a result, all the main directions of thought known to the author of this article are in terms of Finn V.K. no more than games around the base of facts and practically do not operate in any way with knowledge bases representing abstractions as results of abduction.
Just as Michael Giordano’s apt expression that “excessive use of big data is likely to lead to disasters in the field of analysis comparable to the massive collapse of bridges,” the same thing is likely to happen with today's AI efforts.
All of the above is not an affirmation of the futility of neural networks or similar technologies. The scope of tasks and their value are enormous. This and the help in recognition of images and the help to experts in different areas at the analysis of data and details, insignificant at first sight. A good example of such an application is to assist the AI ​​in making diagnoses.
The birth of modern mathematics, of course, is a long process that has been protracted for centuries. Observing current trends, a disappointing conclusion suggests itself: everything moves in a circle. The philosophers of ancient Greece did not know mathematics and mathematical formulas, operated with concepts at the level of images and "everyday" concepts. This was not enough for the organization of more complex and, most importantly, abstract reasoning, which led to the birth of mathematics in its modern sense. Modern reasoning and the logic of what is called "artificial intelligence" today follow the same path. And the state of today "leads" back to basics, since it is more based on the same principles of searching for "common" patterns rather in the style of Pythagoras and Euclid.
One of the main tasks of mathematics is the search for logics, which allows to significantly reduce the cost of calculations by deriving compact and optimal patterns. All this was the impetus for the creation of today's mathematics with its modern notations. The beginning is seen not earlier than the XVI century by a number of scientists such as Descartes, Leibniz, etc. From the point of view of AI - we have not yet reached even the XVII century.
The birth of mathematics capable of induction and abduction is still ahead, and the explosive growth of interest in AI is mainly due to the growth of computational capabilities, and not the emergence of new algorithms. But as a result of this growth, a point was nevertheless reached, after which the solution of a large amount of tasks, both in terms of application and initial data, but relatively small in terms of analytical complexity, became economically viable. But this is still an extensive way of development.
Without solving logical-philosophical issues, without creating new areas of mathematics, in particular, further movement is not possible.
[1] Russell S. Norvig P. “Artificial Intelligence” Williams, 2007.
[2] Osipov G.S. "Methods of artificial intelligence" Fizmatlit. 2011
[3] A. Barsky “Logical Neural Networks” M .: Internet-University of Information Technologies, 2007
[4] Gladkov L. A., Kureichik V. V., Kureichik V. M. Genetic Algorithms: Tutorial. - 2nd ed.- M: Fizmatlit, 2006. - p. 320. - ISBN 5-9221-0510-8.
[5] Pleshakova I.Yu., Chuprina S.I. “Genetic algorithm for improving the quality of semantic search in texts of scientific publications” cyberleninka.ru/article/n/geneticheskiy-algoritm-dlya-uluchsheniya-kachestva-semanticheskogo-poiska-po-tekstam-nauchnyh-publikatsiy
[6] V.K. Ivanov. and Mankin P.I. “Implementation of a genetic algorithm for effective documentary thematic search” Tver State Technical University
[7] Anshakov OM ABOUT ONE APPROACH TO THE GENERATION OF HYPOTHESES IN JSM METHOD
[8] Finn V.K. Artificial Intelligence. Methodology of application, philosophy. KRASAND. 2011
[9] New philosophical encyclopedia: In 4 vols. M .: Thought. Edited by V. Stepin. 2001.
[10] Maturana U. Varela F. The Tree of Knowledge. Prograss-tradition, 2001
[11] Tarasenko V.V. “Fractal semiotics. "Blind spots", twists and turns and recognition "Librokom. 2012.
[12] Pantam H. Reason, Truth, History. Praxis, 2002
[13] V. Tarasenko. “Fractal semiotics. "Blind spots", twists and turns and recognition "Librokom. 2012 p.30
[14] Getmanova A.D. “Logic” “Omega-L” 2015 from 127.
[15] Getmanova A.D. “Logic” “Omega-L” 2015 from 163.
[16] Popper K. Objective knowledge. “Editorial URSS” 2002.
[17] New philosophical encyclopedia
[18] I. Kant. Critique of judgment ability
[19] N. Bostrom “Artificial Intelligence. Stages. Threats Strategies "" Mann, Ivanov and Ferber "2016
[20] Finn V.K. About data mining. News of Artificial Intelligence, â„–3, 2004
Of course, artificial intelligence already exists! If you look at the headlines of articles in popular media, the titles and slogans of various scientific conferences on this topic - certainly it is. It is impossible not to believe, especially when you really want to finally find yourself in the 21st century - the “present”, as described in all science fiction novels. But is it? And if not, then what really exists. In an attempt to understand the myths and realities, this article was written.
Initially, I wanted to start something like this: “for the first time the mention of the term“ Artificial Intelligence ”appeared in D. McCarthy in 1956 at a conference at Dartmund University, the founders of AI should be considered W. McCulloch, W. Pits, F. Rosenblat” and t .d However, it is already too late and does not quite meet the objectives of the article, and Wikipedia is ahead of it with such a beginning.
')
Analyzing the latest "victories" of AI, as well as critical articles, one inevitably comes to the conclusion that everything revolves around several common features. One part of the articles criticizes the impossibility of passing the tests, and the other is full of pathos of "incredible victories." At the same time, the fact that victories were achieved in highly specialized tasks is ignored, where the main advantage of the machine was the speed of searching the base of facts and the “ability” to see patterns where people simply get tired of doing it quickly. Brilliant examples of cluster analysis in one form or another and the formation of a base of template facts. All these are consequences, but the causes in most cases are either not analyzed at all, or are considered superficially.
There is no doubt that today AI can be considered a separate science or, at least, a separate branch of knowledge. Any science must have its own axioms, definitions, theorems and hypotheses. And this is where the fun begins. Namely, from the definition of what kind of activity can be called “intellectual” and what can be described as “artificial intelligence”? Next, it is necessary to designate a list of causes and conditions on the basis of which it is possible or impossible to state the fact of the existence of artificial intelligence in its full form, for example, according to Searle. Uncertainty in these issues and generates the entire huge range of myths that have acquired the AI. In addition, there are a number of quite good articles on the history of AI, for example, here or here . But still, a little immersed in the story once again would be appropriate.
Let's start with the definition of what is AI, let's try to follow the drift of the concept itself. Some dates are approximate.
1956 D. McCarthy suggested that “intelligence within this science means only the computational component of the ability to achieve goals in the world,” and instead of the concept software, computational methods.
1976 Newell-Simon’s hypothesis that the ability to perform symbolic computations is quite sufficient to become able to perform meaningful actions. Therefore, all animals and man in fact perform symbolic calculations. To a certain extent, this is a continuation of the mechanism of Descartes and his Cartesian theory. All modern research is built on this hypothesis, in spite of all its criticism.
1971-79 year. Edward Feigenbaum creates the concept of expert systems, which became the prototype of all expert systems. In fact, there was a turning point, after which the current trend of using statistics as the basis for learning was finally fixed as the main one. Research in the field of heuristics was finally pushed to the sidelines.
1980 year. The first fork. John Searle proposed such concepts as strong (full) and weak AI. “... such a program is not just a model of the mind; in the literal sense of the word, it is itself the mind, in the same sense in which the human mind is the mind. ”
1981 Barr and Feigenbaum determined that artificial intelligence is a field of computer science that develops intelligent computer systems, i.e. systems that have the capabilities that we traditionally associate with the human mind - language understanding, learning, ability to reason, solve problems, etc.
1989-1995 The emergence of the term "intelligent agent". An example is the work of Russell and Norvig [1].
2004 Osipov GS [2] belongs to one of the most accurate, from the point of view of the author, definitions of the term and tasks of AI. He defined AI as a set of methods, models, and software tools that allow artificial devices to implement targeted behavior and reasonable reasoning.
Summarizing all the above, I would like to recall Bertrand Russell's remark: everything that seems definite and precise becomes more uncertain with a closer look. The modern understanding of the term AI in fact increasingly goes to interpreting AI as software or Intellectual agents (AI). The main purpose is to solve a certain class of problems in close intertwining with Big Data as a source of facts for analysis. An example is the translation task. For all its apparent comprehensiveness, it is limited to comparison tasks, but it does not solve the problem of answering the questions: “which mountain is the highest?” Or “what if ...”!
Thus, the problem is clearly observed to drift into the area of ​​highly specialized agent systems and about the actual departure from the task of building an AI-complete model towards specialized intelligent agents. This is not bad news, as it removes the likelihood of a catastrophe described by Nick Bostrom, and the very concept of “intellectual” becomes a mere decoration.
As for the main stages, in due time the chronology which came into our eyes describes well the main stages of the formation of AI, unfortunately the author is unknown:
Stage 1 (50's) (Neuron and Neural Networks)
It is associated with the emergence of the first machines of sequential action, with very small, by today's standards, resource capabilities for memory, speed and classes of tasks to be solved. These were tasks of a purely computational nature, for which solution schemes were known and which can be described in some formal language. The tasks for adaptation also belong to this class.
Stage 2 (60s) (Heuristic Search)
In the "intelligence" of the machine, mechanisms for searching and sorting were added, the simplest operations to summarize information that do not depend on the meaning of the data being processed (in this place it is difficult to fully agree with the author). This became a new starting point in the development and understanding of the tasks of automating human activity.
Stage 3 (70s) (Knowledge submission)
Scientists realized the importance of knowledge (in scope and content) for the synthesis of interesting algorithms for solving problems. This meant knowledge that mathematics could not work with, i.e. Experienced knowledge that is not strictly formal and is usually described in a declarative form. This is knowledge of specialists in various fields of activity, doctors, chemists, researchers, etc. Such knowledge has been called expert knowledge, and accordingly, systems based on expert knowledge have become known as consulting systems or expert systems.
Stage 4 (80s) (Learning Machines)
The fourth stage of development of AI became breakthrough. With the advent of expert systems in the world, a fundamentally new stage of development of intellectual technologies began - the era of intelligent systems - consultants who offered solutions, substantiated them, were capable of learning and development, communicated with a person in his usual, albeit limited, natural language .
Stage 5 (90s) (Automated Machining Centers)
The increasing complexity of communication systems and solved tasks required a qualitatively new level of “intelligence” of supporting software systems, such systems as protection against unauthorized access, information security of resources, protection against attacks, semantic analysis and information search in networks, etc. And the new paradigm of creating promising protection systems of all types has become intelligent systems. They allow you to create flexible environments within which all necessary tasks are solved.
Stage 6 (2000s) (Robotics)
The scope of the robots is quite wide and extends from autonomous lawn mowers and vacuum cleaners to modern models of military and space technology. Models are equipped with a navigation system and all sorts of peripheral sensors.
Stage 7 (year 2008) (Singularity)
Creation of artificial intelligence and self-replicating machines, human integration with computers, or a significant abrupt increase in the capabilities of the human brain due to biotechnology.
According to some forecasts, the technological singularity may already occur around 2030. Proponents of the theory of technological singularity believe that if a fundamentally different from human mind (posthuman) arises, the further fate of civilization cannot be predicted based on human (social) behavior.
Let us return to the problems and myths that have arisen largely due to the large number of writers, as well as in the attempt of different teams of developers to get an investment in their research. In order to expose the myths, not less articles have been written and published. An example is such an article as this or this , but one of the best in terms of figurative presentation of articles, etc. ... To write another article, at first glance, is a rather controversial exercise, if you do not pay attention to some common features. A common feature of all these works is the statement of the fact that all of them are built around criticism of the capabilities of neural networks, as well as the fact that we do not know how the brain works, and that building AIs is not necessarily around the idea of ​​neural models. All this is true, but almost none of the authors are trying to formulate what is needed to solve the AI ​​problem. In the framework of this article, I would like to, albeit conditionally, but try to identify the tasks, without solving which it is impossible to move on.
Before proceeding, I would like to make a few comments.
First, eliminate conversations about clandestine research centers in which some forces hide the true state of affairs. Creating a truly intelligent agent requires solving a number of philosophical problems and developing on the basis of these solutions a new mathematical apparatus, which is simply impossible to hide in full.
Secondly, all further reasoning is by no means a certain statement about the “divinity” or “unknowability” of human nature. The task of this article is to show white spots and show the fact of excessively optimistic statements about the fact of the existence of AI in its full form to date.
What are existing technologies? The total pool of technologies, united under the general term "artificial intelligence systems" is not so great. To him with varying degrees of reliability can be attributed:
- Neural networks
- Genetic algorithms
- JSM methods, etc.
It is possible that it will seem to someone that completely different technologies are dumped here "in one pile". But this impression is deceptive, because despite the obvious differences, they all have common features:
- Assume learning system.
- The basis is a base of facts or a training sample, as a set of samples within the framework of classifying characteristics.
- It is assumed that there are redundant competing calculations until one of the threads reaches a given threshold of confidence.
- The result of the calculation is usually any precedents from a predetermined list.
The very same learning is characterized by the following main features:
- Presence of a priori knowledge given in the form of classifying models
- Availability of a sample base for building a “model of the world” according to classification features.
Neural networks
According to Wikipedia, the link to it will be forgiven, neural networks "are a system of simple processors (artificial neurons) connected and interacting with each other". Handsomely. There are various options for implementation such as Bayesian networks, recurrent networks, etc ... The main model of work: the base of images is the transfer function-optimizer.
The most widely used today are limited Boltzmann machines in a multi-layered version. Layering, i.e. Depth is needed to overcome the “XOR” problem. In addition, as shown by Hinton, an increase in the number of layers of hidden neurons makes it possible to increase the accuracy due to the presence of “intermediate” images with minimal difference in each layer. In this case, the closer the intermediate layer to the exit, the higher the specification of the image.
The main purpose of creating neural networks and the resulting task of learning is to eliminate the need for intermediate computational conclusions when analyzing the profile-matrix of incoming signals. This goal is achieved by creating a base of reference profiles, each of which must have a single neuron at its output - a cell of the resulting matrix. Each such neuron is assigned a certain interpretation-result.
The main problem from the point of view of the problem being solved, as well as the actual training, is the noise of incoming matrices of the incoming neurons entering the analysis matrices. Therefore, one of the main conditions is the availability of high-quality training sample. If the training sample is of poor quality, then a high noise level will lead to a large number of errors. However, the large size of the training set can lead to the same result.
To some extent, the work of the neural network can be compared with the work of unconditioned reflexes of living beings [3] with all the attendant drawbacks.
This approach works well for tasks where there is no strong noise, a clear a priori base of reference images. The main task is to choose the most appropriate image from an existing knowledge base. The tasks of forecasting, in this case, will be solved only by extrapolating the existing history without the possibility of creating new entities, i.e. induction with insufficient deduction.
Some may argue that this is not the case. But before rejecting the above, it is necessary to determine what is meant by the term “new entity”? Is this another instance within the framework of the existing vector-class space of the selected subject area or the creation of new area-spaces? One of the following articles will be devoted to the topic of comparing something with something that can be compared and what cannot.
The basic principle of human learning is induction and abduction. Induction is difficult at least because of the initial formulation of the problem - it will get rid of intermediate calculations. A possible objection to the fact that the processes associated with induction, can occur at the learning stage - are weak, since learning is based on several fundamental principles:
- The presence of the very fact of the emergence of a new profile and its reflection (which is not noise) and the need for an outside expert to determine whether the profile corresponds to the result.
- The absence of a simple and reliable mathematical apparatus that clearly describes the conditions and rules for generating new dimensions, and, therefore, classes of objects.
- By itself, the tracing procedure is a process of generalization, the search for new routes and routes, albeit with the need to control the uniqueness of the correspondence of profiles and results.
- Possible arguments about associative fields do not add anything new, as they are just an extension of the existing deductive approach.
Genetic algorithms
According to the same Wikipedia: “The genetic algorithm is a heuristic search algorithm used to solve optimization and modeling problems by randomly selecting, combining, and varying the desired parameters using mechanisms similar to natural selection in nature.”
There are a lot of works by such authors as Panchenko T.V. [4]., Gladkova L.A., Kureichik V.V. [4], etc ... The basics of the “genetic approach” are well disclosed here, many interesting works on the application of genetic algorithms . For example, the work of I.Yu. Pleshakova, S.I. Chuprina [5], V.K. Ivanov’s article. and Mankina PI [6], articles on Habré and a number of others.
One of the most important advantages of genetic algorithms is the absence of the need for information about the behavior of a function and the insignificant influence of possible discontinuities on optimization processes. As in the case of neural networks, there is a departure from the need to analyze cause-effect relationships, by building a "final" image - the objective function. In this sense, from the point of view of solving text analysis, searching, genetic tasks solve the same problems or are very similar to the methods of latent semantic analysis.At the same time, we must pay tribute, in terms of semantic search and text indexing, genetic algorithms have much greater prospects, compared with the methods of latent-semantic analysis.
From the point of view of pattern recognition, with a very strong stretch, the objective function can be compared with the layer of input neurons and with the expectation of a maximum as an analogue of maximizing the signal of the neuron of the output layer. Although it would be more correct to say that genetic algorithms are used to improve the efficiency of learning of neural networks, they still cannot be considered as a competition for neural networks. Tasks are different.
A common drawback - the absence of induction algorithms - is fully present.
JSM methods
The JSM-method of automatic hypothesis generation, proposed by V.K.Finn (Finn, 1983), is a synthesis of cognitive procedures of induction, analogy and abduction [Finn, 1999]. It can be formulated as a system of rules and even presented as a program in a logic programming language [Vinogradov, 1999], [Vinogradov, 2001], [Efimova et al, 2006]. [7]
The saturation condition will be satisfied if the application of the induction procedure does not lead to the generation of new hypotheses. The test for causal completeness is precisely the use of some variant of abduction. The data set and the variant of the strategy of the JSM method satisfy the condition of causal completeness, if all data on the presence or absence of properties of objects are explained using hypotheses about the possible causes of the presence or absence of properties generated by the induction procedure.
If the condition of causal completeness is not fulfilled, then this is the basis for replenishing the base of facts and / or changing the strategy of the JSM method and / or the way of presenting data. All these operations are performed with the participation of the person. From the point of view of V.K. Finn [8] the most promising in this matter is the further development of JSM methods.
One could add a couple of technologies to this list, but this would not affect the picture as a whole.
Thus, today the term “AI” rather means a subspecies of technological (algorithmic) approaches to solving combinatorial problems. The main tasks of which are reliable separation of “statically significant” in its essence patterns and building images-objects on the basis of statistics, without analyzing cause-effect relationships. The main directions are pattern recognition. Under the images can be understood images, sounds, a combination of symptoms of disease, etc.
The result of learning neural network, etc. there must be some identified pattern, presented in the form of a certain matrix-cluster (vector). Of course, this matrix or set can be constantly corrected by new examples, but this does not affect the essence of what is happening. In any case, the set revealed and cleared of noise can be represented in the form of “alienable logic”, which is a kind of “optimal” way to solve the problem. An example of such an area is the task of automatically categorizing texts, but not from the point of view of posting texts by already known headings, but the creation of headings itself. Their annotation, as well as the automatic construction of various kinds of ontologies.
According to most experts, all these methods are nothing more than reincarnations of the tasks of statistics and clustering. Therefore, all modern efforts in the field of deep machine learning, etc. no more than the efforts of entomologists to study butterflies. Search and description of the color and shape of the wing does not give the slightest understanding of the nature of flight. At the same time, it should be recognized that they are very useful from the point of view of recognizing images of both visual and other similar objects, whether they are medical histories or analysis of chess games ...
All this leads to the conclusion that further reasoning in the direction of research of the physical or "neural" model of the construction of AI is a dead end. All further reasoning must be built on the basis of logical abstractions such as determining what kind of intelligence is, what signs, what “mechanics” of the processes of deduction, induction, abduction. What is the logic and mathematics of the basic laws of logic. What kind of logic and mathematics is necessary, let's say, to “invent” to create AI. What is necessary to solve the AI-complete problem? To begin with, it is necessary to determine the term "intellectual", "intelligence" and what features it should have.
What is considered to be intelligence and even more artificial? And here the first difficulties begin to arise. For example, to consider whether intellectual systems created for solving specific problems, having the potential for logical conclusions and self-learning within the framework of the task field. If, however, we consider an AI as a system capable of self-awareness and goal-setting, can the term “artificial” be considered correct? Indeed, in this case we will deal with a completely independent person, albeit with a different physical nature. Yet “artificial” is not just man-made, but rather limited to artificial borders. Otherwise, in some way all people are man-made, as they are created by other people.
In accordance with the definition, for example, in Ozhegov's dictionary: “INTELLIGENCE, -a, m. Mind (in 1 value), mental ability, mental beginning in a person. Tall and. II adj. intellectual, th, th. Intellectual abilities Intellectual property (protected by the law of someone's work.). " By definition, D. Wechsler "intelligence - is the global ability to act rationally, think rationally and cope well with life circumstances." In Kant (German: Verstand — reason) as the ability to form concepts, and reason (German Vernunft) as the ability to form metaphysical ideas. This usage has become widespread in the subsequent German philosophy and finally settled in Hegel in his concept of reason (I.) and reason. In accordance with the new philosophical encyclopedia "INTELLIGENCE (lat.intellectus - mind, reason, mind) - in a general sense, the ability to think; in epistemology - the ability to mediate, abstract cognition, which includes such functions as comparison, abstraction, the formation of concepts, judgment, inference; opposes the direct types of knowledge - sensual and intuitive; in psychology - rational, subject to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].subordinate to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].subordinate to the laws of logic thinking; opposes the non-rational spheres of the psyche — emotions, imagination, will, etc. ”[9].
Unfortunately, all philosophical definitions, due to their insufficient “mechanism”, do not provide a sufficient basis for solving the problems of constructing AI.
From the point of view of understanding the tasks of AI, the definitions from biology and psychology are closer. For the most part, they boil down to the fact that Intellect is a type of adaptive behavior aimed at achieving a goal. Achievement of goals and adaptation of all researchers is put at the forefront. The analysis of targeting mechanisms is in itself a topic of a separate article, so for now we will limit ourselves to the fact of its presence.
From the point of view of algorithmic problems and logic building, of great interest are the works of Joey Paul Gipford, who in the 1950s proposed a cubic model of the structure of intelligence. Remarkably, world-wide fame brought him research in which he, based on the use of tests and factor analysis, attempted to construct a mathematical model of a creative personality.
The structure of the intellect according to J. P. Gipford.
Separately, A. Turing with his work "Can a Machine Think" and a child machine should be singled out.
The properties of intelligence encountered in different works may already be more specific and understandable from the point of view of algorithms and models for constructing AI. So the list of essential qualities of human intellect, can be quite interesting. These include:
- — . ;
- , ;
- — , , :
- , , ;
- , , ;
- Critical thinking implies the ability to strictly evaluate the results of mental activity, subject them to a critical assessment, reject the wrong decision, refuse to take action if they contradict the requirements of the task:
- the breadth of thinking - the ability to cover the issue as a whole, without losing sight of the source data of the corresponding task, to see the multivariance in solving the problem.
Such properties as “depth”, “proof” and “criticality” are understandable and more or less algorithmized. With the question of the formalization and algorithmization of such properties as “curiosity”, “flexibility and mobility” everything is not so obvious.
To start reasoning about the problems of mathematical logic when creating AI should follow from a formalized presentation of such concepts as “value”, “meaning”, “utility”, “ethical”, etc. ... In fact, there are still no solutions to the problems of formalizing meaning, induction and others complex logical and philosophical questions. For example, what is knowledge, how is it arranged and how is it related to facts. How to calculate the conclusions based on knowledge? What is a sign and what is the minimum unit of meaning? How to operate with predicates of more than first order? Unfortunately, the modern understanding of these concepts inevitably reduces to any functions, clustering, maximizing or minimizing any target parameters. At the same time, differentiability of space is desirable. In reality, thinking processes are not differentiated,rather, they represent something similar to fractals with inno- and self-references. In this sense, for example, the works of Umberto Maturana with his autopoiesis [10], V. Tarasenko [11] or the work of V.F. Turchin, who in 1966 developed the language of recursive functions Refal. Refal was intended to describe languages ​​and different types of processing.
As accurately noted by Pantnam [12]: "concepts ... are not images or mental representations ... but rather abilities and skills that create opportunities for operating with signs for the unity of our behavior and sign systems." And in the language of mathematics, remarkably expressed by Mandelbrot, thinking has an analogy to a fractal. The shape of the fractal does not depend on the sample or the pattern reproduced in the iterations. This form depends on the nature of the infinite iterations of the transformation method [13].
Leaving aside the high matter, we will try to understand at least with such "basic" concepts as "induction", "abduction", "deduction". There are a lot of definitions of all these terms, so we’ll give some comments and try to comment on them. Unfortunately, within the framework of the article it is impossible to consider these issues in detail, otherwise it would have been a book. Therefore, it is necessary to be limited to a superficial overview of some problems. The analysis of such concepts as "knowledge", "fact", "sign", especially from the point of view of algorithms, requires a separate article.
Deduction- inference from the knowledge of a greater degree of community to a new knowledge of a smaller degree of community [14]. From the point of view of algorithmization, perhaps, the most understandable and fully realized concept. All modern rhenium in the field of AI, such as neural networks, just implement this logic.
Induction- inference from knowledge of a lesser degree of community to a new knowledge of a greater degree of community. In mathematical logic, this is a conclusion that gives a probabilistic judgment [15]. The basic problems of induction, formulated by Hume back in 1888 and to some extent solved by Popper K. [16], still / only / already in 1933, were, for example, how to substantiate the belief that the future will be (to a large extent) same as the past? How can one evaluate faith in the fairness of expectations in the future? Is our reasoning justified the transition from cases re-encountered in our experience to other cases with which we have not met before? How to mathematically take into account the habit, as an integral part of the induction process? It would seem that for this there are such things as Bayesian probabilities, etc ... But if you think about it, they absolutely do not give an answer to these questions.It is impossible to make repeatability the main factor in the inductive conclusion and how is the irrationality of behavior? This is only part of the problem. Consequently, a mathematical model based on “classical” statistics does not fully work here ...
The law of absolute quality , formulated by Hegel and populistly paraphrased by Engels as the law of the transition of quantity into quality. In many ways, this law is a refinement of the previous topic. The meaning of the law is that the number of observations is in no way connected with the emergence of new properties. In this case, Engels went to the actual falsification for the sake of his ideas. As well as the problem of induction today has no mathematical solution.
Abduction- cognitive hypothesis acceptance procedure. For the first time clearly highlighted Ch.S. Pierce, who considered abduction (abductive inference) along with induction and deduction. H.S. Pierce believed that by selecting the most essential hypotheses among a vast number of hypotheses, the researchers implement the “abduction instinct”, without which the development of science would be impossible [17]. The problem is that today there is no mathematical apparatus to calculate the threshold, and the rules for accepting hypotheses themselves. How to understand that a certain group of hypotheses explaining groups of facts are “plausible”? Pierce considered the ability of the human brain to such operations "abductive instinct" of man. What is the mathematical model of abduction thinking?
One of the main reasons for the failure of the formalization of thinking processes was formulated by Kant [18] as the problem of the “completion” of artificial systems, in contrast to the natural, which are amorphous in structure and logical “incompleteness”. Thus, all these concepts are quite complex and still not clearly formulated not only from the standpoint of mathematical, but also philosophy as such. Without the implementation of these problems it is impossible to create Intelligent agents capable of solving common problems. But today the author is not aware of this kind of logic. Nick Bostrom’s statement [19] that "... progress on this path requires technological solutions rather than technological breakthroughs ..." seems somewhat more optimistic.
Nevertheless, there are several remarkable studies, such as the works of Finn V.K. in his work “On Data Mining”, [20] and also in [8]. A complete analysis of Finn's ideas would require more than one article, so we restrict ourselves to a brief squeeze.
One of the most important hypotheses, in the author’s opinion, is the logical abstraction proposed by him, which could be called a structural model of any automated intellectual system:
AIIS = (RIS + PIS + ININ) + PAP ,
Where:
AIIS - automated intelligent information system,
RIS = (generator of hepatitis + theorem proving + calculator), Finn V.. called it “Discloser”,
IPR - search and information system,
INIP - intelligent interface (dialogue, graphics, user training in the system)
PAH - subsystem of automatic database replenishment from the “information environment”
AIIS, in turn, has two blocks: a fact database (DB) and a knowledge base (KB). A database consists of objects with a given structure: finite sets, tuples, words, graphs, etc ... The knowledge base consists of: axioms defining the data structure (for example, axioms of Boolean algebra, algebra of tuples ...), connection axioms of the original (for DB) predicates that implicitly define the class of problems to be solved (for example, problems of recognition of causal dependencies).
For the full implementation of full AI according to Surl, there is a lack of understanding where such subsystems as consciousness should be located and the functions that follow from this to determine the purpose of the system or what we call “curiosity”, where to lay the “depth of mind”. Without this, an independent statement of the problem by the system is impossible. We must pay tribute to Finn, in his works he draws attention to this, but within the framework of this abstraction did not consider it necessary to allocate these issues into a separate functional unit.
Analyzing the components, the question arises of the need to solve a huge range of mathematical and analytical problems. With the generation of hypotheses or deduction everything is not so bad, at least there are quite working hypotheses. But as for the philosophical and mathematical questions of induction, abduction is still a complete obscurity. How to work with higher order predicates? What is the general structure of the axioms? How to solve the problems of operating predicates of more than two three orders of magnitude? Create a programming language or operate with something similar to graphs and multidimensional arrays? There is no comprehensible mathematics for work with associative fields. How to analyze the competition of conclusions within the framework of the multidimensionality of space as associative links,so multidimensionality of the space of possible conclusions? Indeed, in each of the “possible worlds” its conclusion will be correct ... What is the logic and mathematics of the problems of inducing quantity and quality? This is only an insignificant part of the issues that are still to be resolved.
As with any scientific hypothesis, the scheme proposed by Finn is not definitive. But today it is perhaps one of the most working hypotheses. The analysis of each of the elements of the scheme itself is worth a separate article.
Studies in the implementation of each of the components, creating a highly efficient bus for exchanging data between these components can be quite promising and exciting tasks. Under the conditions of rapid growth of SaaS, cloud computing, the creation of such systems, built by integrating components from different teams, could give a significant breakthrough in this matter. Intellectual agents created on this principle could really learn to solve a much larger class of problems.
The last remark that I would like to make concerns the term “Artificial Intelligence”. It would be fair to understand the term AI as a general and, rather, a philosophical concept. Within the framework of practical tasks, the use of the term “Intellectual agent” would be more appropriate. This term could provide a clearer understanding of the limitations of the solution in terms of the amount of information, the logic used, the spectra of the tasks being solved and the goal-setting of the systems being created. The use of the word “limitedness” should be understood not as impairment, but as an understanding of the limits of possibilities and to avoid possible substitution of concepts.
Solving the tasks themselves ML, the creation of tensor processors is at best nothing more than digging into one of the components of the Finn model.
findings
Currently, none of the problems of logical inference (deduction-induction-abduction) is solved except for a partial solution of issues related to deductive conclusions. From the point of view of complete AI, not a single problem has been solved at all. The main problem is that the overwhelming majority of studies do not affect the implementation of the abstraction itself, at least in the understanding of Finn, and therefore are deadlocked in terms of AI-full implementation.
As a result, all the main directions of thought known to the author of this article are in terms of Finn V.K. no more than games around the base of facts and practically do not operate in any way with knowledge bases representing abstractions as results of abduction.
Just as Michael Giordano’s apt expression that “excessive use of big data is likely to lead to disasters in the field of analysis comparable to the massive collapse of bridges,” the same thing is likely to happen with today's AI efforts.
All of the above is not an affirmation of the futility of neural networks or similar technologies. The scope of tasks and their value are enormous. This and the help in recognition of images and the help to experts in different areas at the analysis of data and details, insignificant at first sight. A good example of such an application is to assist the AI ​​in making diagnoses.
The birth of modern mathematics, of course, is a long process that has been protracted for centuries. Observing current trends, a disappointing conclusion suggests itself: everything moves in a circle. The philosophers of ancient Greece did not know mathematics and mathematical formulas, operated with concepts at the level of images and "everyday" concepts. This was not enough for the organization of more complex and, most importantly, abstract reasoning, which led to the birth of mathematics in its modern sense. Modern reasoning and the logic of what is called "artificial intelligence" today follow the same path. And the state of today "leads" back to basics, since it is more based on the same principles of searching for "common" patterns rather in the style of Pythagoras and Euclid.
One of the main tasks of mathematics is the search for logics, which allows to significantly reduce the cost of calculations by deriving compact and optimal patterns. All this was the impetus for the creation of today's mathematics with its modern notations. The beginning is seen not earlier than the XVI century by a number of scientists such as Descartes, Leibniz, etc. From the point of view of AI - we have not yet reached even the XVII century.
The birth of mathematics capable of induction and abduction is still ahead, and the explosive growth of interest in AI is mainly due to the growth of computational capabilities, and not the emergence of new algorithms. But as a result of this growth, a point was nevertheless reached, after which the solution of a large amount of tasks, both in terms of application and initial data, but relatively small in terms of analytical complexity, became economically viable. But this is still an extensive way of development.
Without solving logical-philosophical issues, without creating new areas of mathematics, in particular, further movement is not possible.
[1] Russell S. Norvig P. “Artificial Intelligence” Williams, 2007.
[2] Osipov G.S. "Methods of artificial intelligence" Fizmatlit. 2011
[3] A. Barsky “Logical Neural Networks” M .: Internet-University of Information Technologies, 2007
[4] Gladkov L. A., Kureichik V. V., Kureichik V. M. Genetic Algorithms: Tutorial. - 2nd ed.- M: Fizmatlit, 2006. - p. 320. - ISBN 5-9221-0510-8.
[5] Pleshakova I.Yu., Chuprina S.I. “Genetic algorithm for improving the quality of semantic search in texts of scientific publications” cyberleninka.ru/article/n/geneticheskiy-algoritm-dlya-uluchsheniya-kachestva-semanticheskogo-poiska-po-tekstam-nauchnyh-publikatsiy
[6] V.K. Ivanov. and Mankin P.I. “Implementation of a genetic algorithm for effective documentary thematic search” Tver State Technical University
[7] Anshakov OM ABOUT ONE APPROACH TO THE GENERATION OF HYPOTHESES IN JSM METHOD
[8] Finn V.K. Artificial Intelligence. Methodology of application, philosophy. KRASAND. 2011
[9] New philosophical encyclopedia: In 4 vols. M .: Thought. Edited by V. Stepin. 2001.
[10] Maturana U. Varela F. The Tree of Knowledge. Prograss-tradition, 2001
[11] Tarasenko V.V. “Fractal semiotics. "Blind spots", twists and turns and recognition "Librokom. 2012.
[12] Pantam H. Reason, Truth, History. Praxis, 2002
[13] V. Tarasenko. “Fractal semiotics. "Blind spots", twists and turns and recognition "Librokom. 2012 p.30
[14] Getmanova A.D. “Logic” “Omega-L” 2015 from 127.
[15] Getmanova A.D. “Logic” “Omega-L” 2015 from 163.
[16] Popper K. Objective knowledge. “Editorial URSS” 2002.
[17] New philosophical encyclopedia
[18] I. Kant. Critique of judgment ability
[19] N. Bostrom “Artificial Intelligence. Stages. Threats Strategies "" Mann, Ivanov and Ferber "2016
[20] Finn V.K. About data mining. News of Artificial Intelligence, â„–3, 2004
Source: https://habr.com/ru/post/328668/
All Articles