Perspectives of development of a mathematical apparatus of AI or whether there is life outside ML/DL?

Main article: Artificial intelligence (AI, Artificial intelligence, AI)

Evolution of technologies

Natalya Lukashevich, the leading researcher of RCC of MSU, professor of department of theoretical and applied linguistics of MSU, tells:

We see that now vector representations of units of texts of different level prevail: words, offers, documents. It approach began to extend from 90th years of the 20th century when the vector model of information search gained recognition. In this model the document is submitted as a point (vector) in space of coordinates which correspond to the words which are present at a text collection. Now neural networks train vector representations of words on the basis of contexts of their occurrence in texts.

In the field of representation of knowledge, the expert notes, semantic networks of a different type are most of all demanded. So, in hands-off processing of texts examples of such semantic networks are the world famous thesaurus Wordnet and similar resources for other languages (for RuWordNet Russian) and also columns of knowledge (knowledge graphs) which store huge volumes of the structured factual information demanded, for example, in information search, in services of answers to questions, chat-bots. Unlike more difficult logical representations, semantic networks are intuitive and scaled, in them it is possible to implement effective information search and a logical output. Besides, approaches to representation of structures of semantic networks (graphs) in the form of vectors are applied.

It is possible that the combined approaches (neural networks + semantic networks) can be used for improvement of interpretation of the results received on the basis of neural networks noted Natalya Lukashevich.

Achievements of today

In the center of the taking place events there are algorithms of deep learning today. However, in itself they are not the certain "magic wand" capable to give to development of the systems of artificial intelligence unprecedented acceleration at all.

It is just computing structure which for the first time allowed us to perform rather "deep" nonlinear processing of input information at the solution of optimization problems of a certain type. For example, minimization of quantity of errors on selection at supervised learning. Or minimization of the amount of penalties when training with a reinforcement, explains Yury Vizilter, the chief of division of intelligent data analysis and technical sight of State Research and Development Institute of the Aviation systems, professor of RAS.

Quite possibly, the scientist believes that some other too successful computing structure can appear, and it, in turn, will become one more important step forward.

But also it will hardly provide reaching qualitatively new level. Such new revolution is not browsed yet though new solutions need to be looked for always,

There is yet no universal approach which would be comparable to any private approach in its competence, Mikhail Burtsev, the head of the laboratory of the neural systems and deep learning of MIPT, the iPavlov NTI project manager notes. If we want to have a system with the maximum opportunities in each timepoint, for this purpose it is necessary to conduct researches on a combination of the existing approaches for consolidation of their strengths and also to work in parallel on original disruptive theories of AI, he is sure.

Market situation changes very quickly, and temptation to enrich classical predictive models with machine learning is very big. Advanced forecasting, scenario modeling of a situation ("and that if I give a discount for goods A in the amount of the X % as demand for it what share of demand will be delayed on goods substitutes", etc. will change) will actively develop and will have enormous demand. Now does not speak about inclusion of such functionality in the development plan only lazy, notices Sergey Kulikovsky, the CEO of Polymatica company.

By estimates of Sergey Sviridov, the director of research of Tsifra company, the most perspective technology which can find broad application in tasks of industrial enterprises is training with a reinforcement (Reinforcement learning, RL).

However, despite rapid development of this area and prospect which it offers for optimization of production, logistic and other processes its application in problems of the real world bears a number of risks and technology difficulties which slow down implementations in real business, warns the expert.

Solutions of tomorrow

The industry of AI looks for the best combinations of neural networks to other methods. For example, genetic algorithms of different types are used in neural networks and the hybrid systems for optimization of parameters of functions of accessory of input and output variables today. There are works in which the genetic algorithm performs structural synthesis of neural network, in particular, defines type of network, quantity of the buried layers and neurons, selects functions of optimality. In other words, using computer simulation becomes possible to observe development of neural network (or development of cognitive structure) during artificially organized evolution.

As for DL/RL combination to evolutionary and genetic algorithms, it, really, the powerful tool, and it is already used, for example, at the solution of game tasks and AutoML (automatic machine learning). I think, it is just the highway of development of ML in general, considers Yury Vizilter.

Indistinct neural networks

Sergey Germanovich, the technical director of the company Polymatica, sees perspectives in integration of neural networks and indistinct mathematics. However, it is a question not of today. Today, in terms of business, much stabler result is yielded by the scenario modeling executed by traditional methods of numerical analytics.

This assumption, conditionally, it is possible to call correct.

For example, technology which in real time predicts your requirements on the basis of the analysis of set of dynamic data from different sources. What economic result will be received at use of such technology for personalisation of advertizing contacts? It in general cannot be compared to the current "medieval" targeting / retargetingom. Segmentation to within each client already very soon will become a reality, assumes Germanovich.

However, present neuro and indistinct neural networks have a significant restriction - a ratio of accuracy / interpretiruyemosti the received hypotheses.

Neuronets and classical logical methods

Alexey Vyskrebentsev, the head of the center of examination of solutions of Foresight company, compares merits and demerits of classical methods of a logic theory and neural network algorithms. As for machine learning, in the absence of the sufficient volume of observations, algorithms are powerless. Besides, application of neural networks is sensitive to noise level.

In case of logical artificial intelligence the machine itself builds a chain of rules depending on type of events and it is not obligatory that under these rules there were observations earlier. In it the main difference. But also the return is right: if the rule is not registered in logical artificial intelligence, then, unfortunately, without its adding the machine is powerless, emphasizes Vyskrebentsev.

According to the expert, the hybrid techniques integrating both machine learning, and logical artificial intelligence are most perspective:

The first algorithms allow for final time using the machine to create group of rules, – to build the second in a decision making chain, to register exceptions and to find solutions for linear time.

Thus, in the field of mathematical methods of artificial intelligence life boils not really noticeable to an external eye. Active development is followed, first, by consolidation of mathematical methods of different types, and, secondly, also aspiration to a universalization of computing procedures. Time is required in order that in this "boiler" "engines" of highly intellectual processes of new generation prepared. And only after receiving such AI tools researchers will be able to think of how strong can be implemented in practice notorious "are artificial intelligence".

You See Also

• Machine learning (Machine Learning)
• Harmful machine learning (Adversarial Machine Learning, AML)