When the old methods do not work: Vasily Mazin - on a hybrid approach in the development of AI

    Man created the spacecraft and synthetic analogues of living tissues, but was never able to reproduce the mind. What hinders this today: lack of computer power, lack of suitable algorithms ... Or a lack of understanding of what reason is as such? Vasily Mazin, the developer of universal artificial intelligence in the laboratory Mind Simulation, spoke about modern approaches to the development of AI, methods of teaching virtual personalities and a little about the philosophical aspects of mind simulation.

    Vasily Mazin , director of research at the laboratory Mind Simulation.

    Interviewer - Elena Andreeva

    On the Mind Simulation website , it is reported that you now have two projects: Core Expert, universal artificial intelligence, and Steve, an assistant with AI. Tell us more about them: goals, audience, stage of development?

    Core Expert is a universal intellectual core with a modular structure that can be used for various AI solutions. This is the main product of our laboratory. The ultimate goal of its development is the creation of universal artificial intelligence, the scope of which is limitless.

    Designing Core Expert took about 10 years. We carried out theoretical development, algorithmic design of the main modules, and now for about a year we are in the implementation stage.

    Steve is a kernel-based product: it makes the most of all the components in Core Expert. In fact, this is a demonstration of the capabilities of the system, which we plan to turn into a full-fledged intellectual assistant in the future. Now Steve is at an early stage of development, but is already in the public domain.

    What devices is Steve available for?

    While there is only a browser version that is available on the site, we have not done the application yet, although it is in the plans. In the same place you can observe how it grows and develops: every week we release news describing new features and capabilities.

    Tell us about the audience of these projects. Is your segment B2B or B2C?

    Core Expert is universal, we can build systems based on this kernel for solving various classes of tasks. And for Steve, we see two directions. The first is for people: to solve personal problems and tasks. And the second is for business when it is used internally and helps all employees in their work.

    In the description of the mission of the laboratory Mind Simulation states that its product, universal artificial intelligence, will not be trained in the dialogues of people. What databases do you plan to use and from which source?

    Nowadays, the developers of most popular voice assistants use large corpuses of human language dialogues. Programs trained in such bases work without knowledge and understanding of what is at stake. Often, this leads to funny and frightening situations: there are many collections of users on the Internet with strange answers from “Alice” and “Alexa”. We tried different approaches and learning strategies. Considered various open databases and anthologies that exist. And they did not suit us, because the information there was either incomplete or not in the form in which we need it. Therefore, we have come to a hybrid learning model. The system has been conducting a dialogue with users for the third week and identifying gaps in knowledge. These spaces are used to form the vector of the request in a special program for obtaining knowledge.

    The knowledge acquisition program works with a large corpus of texts. It uses well-known linguistic and statistical methods for studying information; formed a set of copies of knowledge. And the final decision on the addition is made by our experts.

    It is assumed that the system will be functions of learning and self-learning; the foundations have been laid now. In the first stages, the tool will be inference: for example, induction and deduction, analogies, fuzzy logic and imitation. In the future, we will endow him with curiosity and fantasy. We are moving in stages, so I can’t tell you exactly how “machine curiosity” will be realized: when we get to this stage, there will be a topic for a separate publication.

    Tell me, how do you use other disciplines in the development of your artificial intelligence?

    Core Expert thinking does not depend on the language of communication. We use linguistics in an auxiliary module, for learning: to translate knowledge from a natural language into an internal representation of the nucleus and vice versa. So far, we apply psychology only to an intellectual assistant: there we have the task of modeling a unique personality. In particular, we considered various personality models and the way they are reflected in the dialogue module, and now we are continuing to work on the formation of the virtual personality of our assistant.

    Why is the hybrid approach called the future for artificial intelligence systems?

    Now there are about 20 different technologies for AI solutions. At the same time, over the past 20-30 years no new technologies and programming paradigms have appeared, and there are big doubts that they will appear at all. Therefore, for further advancement, it is now necessary to combine existing approaches. Recently, more and more researchers are conducting experiments in this direction and are achieving quite good results.

    Tell us more about the device hybrid systems.

    Hybrid systems combine several classic approaches to solving problems in the development of artificial intelligence. I divide them into two categories. The first is “weak hybrids”: when, within the framework of one task, separate technologies are used together and exchange data through a special interface. This may be the determination of the main components of neural networks by methods of genetic algorithms or the joint use of neural networks of different architectures. The second category is “strong hybrids”: when two models are so deeply embedded in each other that a new technology actually appears. Ideally, it inherits the strengths of each model. For example, it can be fuzzy neural networks, they combine a fuzzy logic apparatus and neural networks.

    What do you think, what technical limitations today hinder the development of AI?

    It seems to me that now there are no technical limitations. The available capacity is enough to solve any intellectual problems, the only question is how to solve them. All the pieces of the puzzle we have, it remains only to put them correctly. But the emergence of a full-fledged computer mind - this is a more complicated question. It touches upon the concept of mind as such and the question of things necessary for the manifestations of its essence: body, emotions, feelings, and so on. This is not only an interdisciplinary task, but also a very philosophical one, because we are not yet completely clear what reason is.

    Tell us about your presentation at AI Conference . To whom is it addressed and what topics will it reveal?

    I will talk about our approach to creating an intellectual core, describe the main components and how they interact. I will tell you about the advantages of this approach, the prospects for application - including for solving problems in various areas of business.

    The report contains specific terminology from the field of AI and is designed for basic knowledge in this area. In addition, throughout the day, we will have an open Steve demo booth and any visitor will be able to chat with him and our developers, ask them questions and evaluate the current capabilities of our products.

    We plan to pay more attention to Core Expert in the report, and focus on Steve at the stand.

    Also popular now: