Are artificial intelligence systems today actually more intelligent than those designed 20 years ago?
And how to determine whether robots or systems designed to extract information from digital libraries are actually becoming more intelligent?
The answers to these questions are crucial, among other situations, to determine whether there are real gains to update the version of an artificial intelligence system.
And in the future, may represent a metric for assessing whether a new generation of robot assistants is actually more intelligent than the old robot that has helped him so many years.
Universal Intelligence Test
To help answer such questions, researchers look for what they call the Universal Intelligence Test.
According to the theorists, there are several essential elements to characterize an intelligence test truly universal.
Everything starts with a test that can be applied to any subject, be it biological or not.
The second requirement is that it can be applied to the subject at any stage of their development. If a biological being, the test should also work for children and adults, if a non-biological, the test should be applicable to any version of "mechanism".
It should also be ensured that the test does not become obsolete: to be universal, it should be used for any subject or system, now and in the future.
The universal test of intelligence should not be prejudiced: he must be able to measure any level of intelligence on a scale from zero to genius.
Finally, the test is incremental and conclusive on any point in your application - in other words, it should be possible to stop the test at any time and still have been a reasonable measure of intelligence.
Calculation of complexity
This characteristic of a test that can be interrupted at any moment is particularly challenging because it is very different from the current psychometric tests used by psychologists, or even tests of intelligence, as the Turing test.
But it is not something impossible to do. At least that is what is concluded in a paper presented by José Hernández-Orallo, Polytechnic University of Valencia (Spain) and David Dowe, Monash University (Australia).
"We have developed an intelligence test that can be interrupted at any time, but it gives a more precise idea of the intelligence of the subject tested if there is more time available to accomplish it," say the researchers.
They used interactive exercises in environments with a difficulty level estimated by calculating the so-called "Kolmogorov Complexity", an indicator that measures the number of computational resources needed to describe an object or a block of information.
The use of mathematical and computational tools for structuring the test makes it conceptually very different from traditional tests that result in intelligence quotient.
Human paradigm
According to the researchers, this test is the first approach to something that allows you to systematically assess the progress of artificial intelligence systems, checking, for example, if indeed there is more intelligence built into an existing system than in a previous version or a system 20 years ago.
Even to the theory, however, still failed to overcome the "human paradigm": there is no theoretical concept that trace the course to assess whether a system would have more intelligence than human intelligence.
"The universal and unified assessment of intelligence, whether human, animal non-human, alien or artificial, was not being approached from a scientific point of view so far, and this [our work] is a first step," concludes researchers.
Nenhum comentário:
Postar um comentário
Comente