Where did neural networks come from and what is happening now

    In the past few years, the topic of artificial intelligence has been actively discussed, since one of the approaches to its study is actively gaining momentum among large corporations. This approach is a neural network. Not long ago, about a year ago, this word could be heard from everywhere. Today we will look at the history of the study of artificial intelligence by mankind (it turns out that it is already about 2000 years old) and today's realities.

    The authors of this article are guys from Jedium. Give them the word.


    And indeed, thanks to the growth of computing power of ordinary computers, it became convenient to engage AI at the program level. Up to the fact that the popular IT sources laid out articles on how to "write" your neural network with minimal programming knowledge. However, before delving into the study of our company in this area, we would like to tell a little background.

    AI is not a 21st century discovery.

    The rapid development of this industry began several years ago, but before that, mankind has been studying artificial intelligence for a couple of thousand years. Starting from Aristotle and Descartes and ending with the notorious John von Neumann. The latter made a huge contribution to the development of the logic of a modern computer. There are very many scientists who have studied this area. On a full description of the history of AI will take more than one series of articles with long transcripts and explanations. The purpose of the prehistory of the science of AI is to convey several theses that will be important for understanding what Jedium does.

    In the 1930s, there was a booming interest in the study of artificial intelligence. Scientists with different approaches to the study of AI achieved excellent results, challenged each other's theories and proved new ones. For example, at this time, the Turing test was invented, suggesting that if a person cannot distinguish AI from another person in a conversation with two interlocutors (the conversation is supposed to be conducted using a computer terminal to eliminate the influence of voice, appearance and similar discrediting qualities). ), then we have a full-fledged AI. The most interesting thing is that once the test was passed. As a result, Turing's work became the fundamental base of knowledge about computing systems, which, on the basis of input data, create their axioms (judgments that do not require proof). Thanks to the axioms, Systems can draw conclusions on specific requests from the user or build predictions with very impressive accuracy. Nowadays, these systems became known as neural networks.

    There are several approaches to the study of AI, but it is worth noting at least two - with certain semantics and without it. Either we write logic, or we write a high-loaded computing program, that is, a neural network. However, after the end of the 40s of the last century and to the present, no major breakthroughs were observed in the AI ​​sphere. It was “winter” - a decline in interest in the study of artificial intelligence. This was due to the lack of computing power to build powerful neural networks, and the approach from the logical side turned out to be overwhelmingly difficult and showed disappointing results.

    A few years ago, the computing power was enough to design very powerful neural networks. However, the interest of ordinary people in them is acquiring a negative trend, while the opposite is true for specialists - training continues. Even it is worth making a reservation - more and more neural networks are beginning to acquire the property of everyday and everyday things. Voice assistants in your gadgets, machine vision, which you once again use to scan documents with your smartphone, photo editor and many other examples. People get used to it and cause a “wow effect” more and more difficult. Therefore, there is a thesis that the development of neural networks will not move as quickly during the near future, in other words, a new “winter” will come.

    But even though the work on the study of AI in corporations does not stop, the industry has serious problems that hinder productive research in this area. For example, the lack of data for training neural networks. Studying this conclusion and asking the question “where can we get the data from?”, We found an interesting approach that can be a solution to the problem.

    Living example

    The guys from Jedium are creating a platform that standardizes and simplifies the development of VR / AR applications, as well as creating learning systems. Working in this area, they found interesting research on so-called “hidden knowledge”. Hidden knowledge is the skills that one person has, but others lack. A person with hidden knowledge can share them. Only the whole problem lies in the name - these skills are hidden, and until a person is told about them, he does not know that he has them.

    Hidden knowledge is the basis of social learning. If there is knowledge that one student can give to other students, then the burden on the teacher is reduced. Practically no one has yet tested this thesis in practice - there are no specialized software systems. The company is moving in this direction. There is a virtual environment in which you can recreate the conditions for recording hidden knowledge, and then use it to transfer them to other students.

    But to take advantage of hidden knowledge, and simply to create an effective learning environment, we need fairly strong elements of artificial intelligence. For example, to build a truly individualized learning (Tailored Learning), based on the gaps in knowledge and preferences of each of them. And for this, effective data analysis algorithms are required, in the modern reality - trained neural networks.

    Modern e-learning

    Creating a modern online learning system, you should also take into account modern trends, and there are several of them, and it is often difficult to distinguish the line between them:

    • The transition from learning “with the teacher in the center” to learning “with the student in the center”. For quite a long time in the online learning systems, only the classical paradigm was considered - “the teacher tells the students something”. At the same time, it is not very important what specific technical tools were used, whether it be modern LMS and tools for creating educational content or just presentations sent to students. Now there is a general opinion that the focus of the learning system should not be a teacher, but a student who receives knowledge from various sources and forms his own picture of the world;
    • Asynchronous / synchronous communication. This is a somewhat smaller problem in the technical sense, as there is now no shortage of tools for live remote communication. A much greater problem seems to us to be the proper use of such means, especially in the context of planning for learning in general. Despite the fact that the concept of blended learning is not at all new, we quite often faced situations where live communication and the training course itself were separate from each other, without being combined into a single system. If we talk about methods of social learning, they generally developed for quite a long time “outside the mainstream”, which, of course, created a number of fairly interesting products such as the Knowledge Forum, but left open the question of how to combine such approaches with the generally accepted ones.
    • Constructivist and connectionist paradigms. Again, both of them look very interesting in combination with all of the above, but we believe that no common patterns of their use and implementation in a software product have yet been found (despite the fact that for traditional LMS / LCMS all this has long existed).
    • Simulations, serious games, role-playing games. In many areas, these types of trainings are considered the best. But at the same time, for a long time, they were somewhat aloof from the training systems in general - neither standards nor application practice contributed to this. Now, with the transition from AICC / SCORM to xAPI, there is a clearly visible opportunity to integrate them into training, but, again, there are no patterns or best practices.

    Creating our platform, we looked for approaches to solving some of these problems, while trying to find a solution not in theory, but in a very specific software product with certain capabilities. We believe that this was partially possible, but we also see that a huge part of these problems requires further development of the platform and analysis of its work. We would like to talk about this in more detail in future articles.

    Expanding the limits of intelligence

    Our mission is to expand the boundaries of intellectual development, enriching the social knowledge of people with technological advances.

    The theory of teaching people has long developed in isolation from technological progress. Our goal is to erase this barrier by achieving synergies between human and machine intelligence.

    And to understand in the process what is actually happening.

    The authors

    Company Jedium - Microsoft Partner company working in the field of virtual, augmented reality and artificial intelligence. Jedium has developed a framework to simplify the development of complex projects on Unity, part of which is publicly available on GitHub . Jedium plans to replenish the repository with new modules of the framework, as well as integration solutions with Microsoft Azure.

    Vitaly Chaschin - Software developer with more than 10 years of experience in the design and implementation of three-dimensional client-server applications - from concept to full implementation and integration of applications and solutions in the field of virtual reality. Systems Architect Jedium LLC, MSc in IT.

    Alexey Sarafanov

    marketing manager at Jedium LLC.

    Sergey Kudryavtsev

    CEO and founder of Jedium LLC.

    Also popular now: