In this review article, I talk about my experience in developing artificial intelligence (recreating the human psyche), what results have been achieved (basic cognitive processes of a person, understands the text and can ask clarifying questions), and in what solutions is the technology applicable at this stage of development (AI is ready replace live online consultants and can be a virtual assistant for programmers).

Understanding in the context of AI

Understanding is one of the important abilities of the intellect. The importance of the concept of understanding for artificial intelligence can be traced to the ideas of Alan Turing, Marvin Minsky and Ray Kurzweil.

According to Wikipedia, “understanding is a universal operation of thinking associated with the assimilation of new content, its inclusion in the system of established ideas and ideas” . Key in this formulation, we believe the need for the AI ​​"system of established ideas and ideas." In order for an AI to understand how a person, the system of knowledge of AI must be identical or very close to the system of human knowledge. Otherwise, the person will be perceived, but will not be understood.
')
On the practical side, the realization of understanding in AI makes it possible:

- to teach AI an array of content created by a person for a person;
- interact with AI in natural language;
- receive conclusions from AI, which are based on the understanding “like a person’s” (the answer “42” is not always what we need).

Therefore, we have adopted the concept of Understanding, as the main one in our approach to the development of AI.

The main approaches to the development of AI - "top-down" and "bottom-up"

In the development of AI, there are two main approaches. The first is aimed at reproducing in a computer the cognitive abilities of a person without reference to the level of individual neurons. This approach is called top-down. The second approach is aimed at building intelligence from neurons to general levels of cognitive processes, and is called bottom-up, respectively.

As for the general trends in the development of AI, the bottom-up approach now prevails. Probably the mathematical apparatus used in artificial neural networks is more understandable to scientists, developers, and enthusiasts.

We use the first approach - “top-down”. Usually this direction of development of AI is complicated by the lack of a common, more or less coherent, consistent and comprehensive theory of human mental processes.

I put together a team to work on such a theory in 2003, then for purposes not related to artificial intelligence or to information technology in general. A group of academic and practicing psychologists and psychotherapists set themselves the task of streamlining the theoretical concepts and practical methods available in psychology. I had to critically review all the available scientific knowledge on this topic. More than 30 people took part in the project at various stages.

By 2008, we have developed a so-called. model of the psyche - a consistent concept of the organization of the psyche and the processes in it. The model turned out to be complicated due to the large number of interacting elements, but it was also a plus - it was well algorithmized. Those. the model operates with some elements and describes the nature of their interaction according to clear mathematical laws.

The development results were tested in several product solutions. In particular, we modeled the behavior of Internet users when choosing a news article, choosing a result from the search results, and also divided users into groups of consumers of different types of goods based on simulated needs. As input for the simulation, we used the history of each individual user visiting the sites. By the way, these data are enough to calculate about a hundred characteristics of a person with which the psychologist operates - extroversion, compulsiveness, etc. Gender is also a psychological characteristic, but not for all countries. For example, for Sweden, the determination accuracy is at most 65%, while for Eastern European countries, we can determine the gender of a user based on the browsing history with an accuracy of 95%.

We understand well what processes are taking place in the psyche, we have the tools that allow these processes to be researched and analyzed. We use the phenomenological method (1). This method allows you to select the processes of the psyche, to share them among themselves, to observe the dynamics. For us, the psyche is not a black box, for the most part it is a complex but understandable structure.

Also note that the approach of our team implies the universality of the developed AI, in contrast to specialized AI, focused on solving single-type tasks.

Storage and processing of knowledge in AI

The first task we have begun to solve is the development of a semantic network for storing knowledge as it is stored in the human psyche. At the same time, the structure of the semantic network should ensure not only data storage, but also the efficient operation of algorithms that repeat cognitive and other processes.

The vertices of the network are any phenomena that consciousness can distinguish. Vertices are of different types, for example, an object, an action, a sign, an abstract concept. Examples of peaks - laptop, swim, black, space.

The connections between the peaks reflect the type of interaction of these phenomena in the psyche. In the expression “blue ball” between two peaks “blue” and “ball” there is a characterological connection. Links can also be of several kinds.

In their work on the formation of semantic network structures, i.e. to add new types of vertices and connections, we follow the process of ontogenesis of human cognitive functions. Let us clarify this with an example. Up to a certain point, there is no “instrumental” connection between the phenomena in the psyche, because the psyche does not solve the corresponding tasks - the child may knock the spatula on the toy, while he does not perceive the spatula as a tool. After a while, imitating what he saw and accumulating knowledge, the child formed in the psyche a new kind of communication - instrumental, the child realizes that you can dig with a spatula. Thus, the psyche responds to the complication of operations, to increase the amount of data, and optimizes its work by forming a new type of communication.

Formation of new types of communication occurs in connection with a new type of activity and in connection with the appearance of neoplasms in the psyche. (2) An example of a new type of activity is that a professional designer has color not only a characteristic, but also a professional tool that forms a separate type of connection. An example of a neoplasm in the psyche is abstract thinking.

First of all, we have identified two stages in the formation of the psyche — preverbal, when structures are formed to denote which words are not used (from birth to 2 years), and verbal, when the word is used as an identifier. The division into the indicated stages is conditional, since each of them also has its own separate periods of the formation of structures.

Reproduction of preverbal structures turned out to be quite a challenge. We were forced to reconstruct concepts and connections from structures known to us from later stages. For example, the pleasure of a six-month-old child does not yet contain differentiated experiences (be it sexual pleasure, pleasure from satiation, pleasure from intimacy, etc.). In the study of preverbal structures, we have limited ourselves to those that have connections with the structures formed in the subsequent stages. Note that although the method of forming preverbal structures for the semantic network has been worked out, this task requires a significant amount of work by psychologists. At this stage, we have identified 3 types of vertices and 4 types of links.

As soon as the word became the main identifier of information, it became easier to form structures. You can always refer to the original source - the actual psyche, asking questions of a person of the appropriate age. Note that at the verbal stage, the psyche is not limited to an extensive increase in the number of phenomena. There are also new types of phenomena, such as abstract, invented, etc., which entails a change in the structure of the network.

To date, there are a number of developments seeking to repeat the principle of storing knowledge by a person using a semantic network. Our method of forming a network structure is based on the tools we have developed for studying mental (including cognitive) processes. We brought into the network structure only what is actually present in the psyche of the corresponding age. As a result, the semantic network, which recreates the cognitive structure of the psyche of a seven-year-old child, contains 17 types of vertices and 15 types of communication. By this age, most types of communication have already been formed in humans.

In the process of creating a semantic network, we identified several interesting features:

1. The semantic network has some tricks - the elements that determine the uneven distribution of links. Most of these chains are oriented around them. This is a human need.

2. With each neoplasm, the structure of knowledge is rebuilt. The new age period brings the need to significantly change the structure of relations between the elements.

3. Creating a certain reference network, we fixed personal differences. These differences are typical. That is, the number of invariants is very limited and due to individual features or the influence of the environment. These are psychological types.

Getting new knowledge and forming answers to questions

Algorithms that implement the operations of thinking associated with the understanding of new knowledge, based on the structure of the semantic network.

First, the text is processed by a parser developed at Stanford University . Then we check the availability of the corresponding vertices and connections and, if necessary, form new ones. When forming the answer or the clarifying question, the algorithms are guided by the available data in the semantic network. Simplified operation of the algorithms can be demonstrated by the following example. Having read the text “Polar bears hunt for seals”, the AI, let us say already knowing each of these phenomena separately, builds a special connection in the semantic network that captures the exclusivity - it is the polar bears who hunt for seals. After receiving the question “Brown bears hunt for seals?” AI will answer “No”, because there will not be a structure in the semantic network that meets all the specified conditions “Brown bear hunts for seals”.

At the first stage, we developed algorithms that correspond to the cognitive processes of a three-year-old child. On the basis of algorithms, we wrote a program that can understand a very simple text and answer several types of questions, here is a demo program . The program allowed us to customize the connections of the semantic network on large amounts of data and confirmed the direct correspondence of the response of the system and the psyche of a living person.

In September, we completed a simulation of the cognitive processes of a seven-year-old child. As long as we do not have the resources to program the entire system, we are therefore temporarily limited to the semantic network and several types of algorithms. If the algorithms at this level can be calculated “manually”, then this is not the case with the semantic network - it contains about 10,000 vertices and 40,000 links. After we finish programming all the algorithms, this solution will be a powerful enough cognitive system for working with textual information. Our system can provide an understanding of new knowledge described in natural language, and operating on the same level as that of a person of a corresponding age.

Next development steps

The next step in the development of our technology we see in the reconstruction of the abstract thinking of man. Now AI can operate with abstract concepts that it has been taught, but cannot independently form new ones. For example, here is the text used for teaching the concept of fish:

Fish is an animal. Fish has elongated body. Fish has flattened body. Fish has head, jaws, gills, tail, silver scales. Fish lives in water. Fish can swim, sleep, eat, feel pain, fear. Fish does not speak. If fish is out of water then it dies. Fish uses gills for breathing. Fish uses fins for swimming. Salmon, trout, shark are fish. Cat, bear, coyote, alligator, seal, pelican hunt fish.

Independent formation of abstract concepts will significantly expand the possibilities for teaching AI and the area of ​​its use. We have already developed basic principles and in the near future we plan to detail the processes. Adding abstract thinking will also require upgrading the semantic network because new types of communication will be added.

In addition, we plan to develop algorithms for the independent formation of new types of communication by artificial intelligence. This will give the plasticity of the semantic network and some "autonomy" in the training of AI from a team of experts.

Application of AI technology

The developed solution is capable of understanding plain text, for example, most articles from Simple English Wikipedia can be understood by the system. Also, our solution can answer questions about existing knowledge and ask clarifying questions if it finds a contradiction. Of the limitations at this stage - texts and questions should be constructed grammatically correctly. In addition, the system should be trained by our experts in abstract concepts in a new subject area. With this approach to learning, the developed solution is indistinguishable in dialogue from a real person.

Of the features of our approach to learning, there is no need for a large amount of learning data. New abstract concepts and new knowledge in a particular subject area, the system is trained by reading the text once.

The cognitive processes recreated in our AI are already enough for use in smart bots, online text support, NPC in games, etc. Those. in systems that require communication in a separate subject area in a natural language. Now the answers in such decisions are formed by live operators or scripts. Our system forms the answer from the existing knowledge in the same way as a person.

We can add certain types of cognitive algorithms and teach the relevant subject area that the AI ​​could perform, for example, entry-level programming tasks. This is a virtual assistant who will take on some of the routine programming tasks - he will understand the task in natural language, analyze the existing code, write a new one. First of all, we are talking about algorithmic problems with a good formulation and with a low degree of uncertainty.

Own resources are not enough, so we are considering a partnership with other companies. We can fully provide the technological part for products or platforms, and from the partners we expect to see product expertise and resources for implementation in the code. Also, the team will be glad to developers who are close to the ideas outlined in the article.

In the comments, please write about what aspects of technology you would like to learn more, in the following articles I will try to take this into account.

1. Husserl E. Logical studies. / Trans. with him. E. A. Bernstein, ed. S.L. Frank. The new edition of R. A. Gromov. - M .: Academic project, 2011.
2. Piaget J. The speech and thinking of the child. - M., 1994.