- Статьи
- Science and technology
- In short: chatbots and robots will be taught to work without the Internet
In short: chatbots and robots will be taught to work without the Internet
Russian scientists have proposed a solution to a complex mathematical problem formulated about 30 years ago. It involves recovering large amounts of data from a limited set of key rows and columns. The result can be used in machine learning technologies, including chatbots, recommendation systems, and search services. In addition, the developed approach opens up opportunities for creating more compact neural networks capable of operating independently, without a permanent Internet connection. According to experts, the implementation of such solutions in real products will require additional work on the reliability, stability and security of autonomous AI systems.
How to optimize work with large databases
Scientists from the AI Institute of Artificial Intelligence have proposed a way to speed up machine learning systems, including chatbots, recommendation and search services. It is based on solving a mathematical problem of restoring large data tables with minimal errors in key rows and columns. Previously, this was proved only for the special case of 2×4 matrices. The new result significantly expands the scope of the approach, extending it to variants with an unlimited number of lines.
A complex mathematical problem from the field of matrix theory was formulated in 1997 by the Russian mathematician, academician Evgeny Tyrtyshnikov. It assumes that any large table of numbers (matrix) can be compressed by selecting a "skeleton" from it — several of the most important rows and columns, and then using them to restore the original data. According to the hypothesis, for a special class of matrices, the error grows very slowly and almost does not depend on the size of the table.
— At first glance, this is a purely abstract task. However, the modern world is experiencing explosive growth in data processing volumes. The result shows that for a wide range of tasks, it is possible to select data more reliably and cheaper than previously thought. This provides a more solid theoretical basis for data compression and approximation algorithms," Mikhail Pautov, a senior researcher at the AIRI Institute, told Izvestia.
According to him, the implementation of the found principle allows developers of artificial intelligence systems to obtain mathematical guarantees that large models can be significantly reduced without drastically reducing the quality of work. In addition, the task of identifying a limited number of the most informative users and objects is simplified while maintaining the accuracy of the output.
Similar approaches are already being used in image, video and scientific data processing, where the key task is to reduce computational costs without significant loss of information, the scientist said.
How AI assistants and robots can do without servers
— The main value of such research is that mathematics teaches us to find order where a person sees only chaos. The more information humanity produces, the more important the methods of sampling essential details among a huge number of secondary ones become," said Richik Sengupta, co—author of the study and senior researcher at the AIRI Institute.
Among other things, the solution will allow creating compact neural networks that, like Google's Gemini Nano, will be able to work on personal devices without connecting to remote cloud servers, the researchers explained. According to them, such AI assistants will be able to perform basic tasks — answer questions, translate texts or suggest recipes — even without Internet access. This will be possible due to the fact that simplified neural network frameworks hosted in the cloud can be stored directly on smartphones.
Other uses include streaming services that select movies and music based on a small set of user profiles. The found principle will also help in the creation of autonomous drones and robots that do not need a permanent connection to the server for navigation and decision-making.
— The problem of computing power is one of the key issues for AI systems, as the volume of data, the size of models and the complexity of their architectures are constantly growing. To solve the problem, companies are developing infrastructure and optimizing algorithms. At the same time, part of the calculations can be performed on users' devices — smartphones and laptops. This uses additional resources and reduces the latency of services," said Georgy Yakushev, a researcher at Yandex Research.
Despite the limited performance of portable devices, even a slight acceleration of their operation can significantly affect the development of AI products, the expert added.
In his opinion, the main difficulty of implementing autonomous AI systems is not so much the quality of data generation as reliability and security, since such agents must correctly understand the task, be able to work with tools, check the result and not commit undesirable actions. Therefore, the development of such applications, in addition to strong models, requires control mechanisms, transparency in decision-making, and deep integration with infrastructure and services.
Skeletal approximations, which are ways to represent huge matrices in a compact form, are especially interesting because they use real reference objects rather than abstract "hidden features" (numerical characteristics that are understandable to the algorithm, but usually do not make clear sense to humans). For example, characteristic users and products in the recommendation system or the most significant features in tabular data," added Egor Samokhvat, head of Data Science at Avito's monetization department.
According to him, the presented solution can improve the efficiency of the search for reference objects and make approximate calculations more stable. The next stage of the scientists' work will be the transition from the proven case of matrices with two columns to the general form of the problem. This step will bring the theoretical results closer to the typical practical tasks of modern machine learning systems.
Переведено сервисом «Яндекс Переводчик»
