Israeli mathematicians have proven that artificial intelligence is far from always able to find patterns in datasets or give unambiguous answers to any questions. Their findings were presented in the journal Nature Machine Intelligence.
Modern machine learning and artificial intelligence systems operate on a very simple principle. They gradually learn to "see" certain patterns and to distinguish correct answers from incorrect ones using extensive human-prepared databases.
Initially, this approach was used mainly to create image recognition systems. Subsequently, it turned out that it can be used for almost everything, from “creative” AIs, able to draw and create music on their own, to the AlphaZero machine, which can learn without the help of people and play several board games, knowing only their rules.
Such successes, Yehudayoff notes, have forced programmers, philosophers and mathematicians to wonder whether this method of problem solving has limits and whether an extremely "general" artificial intelligence can find a pattern in any arbitrary set of data and answer all possible questions.
Israeli mathematicians tried to find out if this is really so by analyzing the most general versions of various mathematical problems that are actively solved today using machine learning systems.
Their attention has been drawn to versions of artificial intelligence that try to predict maximum values using incomplete datasets. For example, such machines try to guess the preferences of visitors to a particular site and select such ads that would be interesting for most of them.
By presenting this problem as a collection of several large and small sets, Yehudaioff and his colleagues found that it was similar in its description to the famous Gödel theorem. Back in 1940, the famous Austrian mathematician Kurt Gödel found out that any formal system, including mathematics itself, is incomplete or contradictory.
In other words, this means that for machine learning systems, as well as for “simple” mathematicians, there are problems, statements and questions that can neither be solved, nor proven, nor disproved without going beyond them.
Promotional video:
In this case, for example, it is impossible to predict whether artificial intelligence can be “trained” to ideally match ads using knowledge of the preferences of only a small, arbitrary number of visitors. Depending on which portal visitors will be included in this sample, this problem is both solvable and unsolvable.
As scientists emphasize, from a practical point of view, this discovery does not affect in any way how actively artificial intelligence will develop in the future and how well it will solve practical problems. On the other hand, the presence of such restrictions suggests that it will be much more difficult to create a universal "thinking" machine capable of solving any problems than scientists believe today.