DeepMind researchers tested the abstract reasoning ability of artificial intelligence. Although artificial intelligence is far behind people in abstract reasoning, the results are promising. Artificial intelligence can successfully perform the tasks assigned to it in different fields of application today; it can even get ahead of people at some point. In contrast, artificial intelligence is still quite primitive in abstract reasoning. DeepMind, the artificial intelligence company that Google acquired in 2014, developed a new test method to see how artificial intelligence is at the level of abstract reasoning and how it can address its shortcomings.
Artificial intelligence-specific testing DeepMind states that seeing beyond the current situation and being able to relate concepts is the basis of abstract reasoning; He underlines that many discoveries, which are the basis of science, are created by abstract reasoning, as Archimedes finds that the volume of an object is equivalent to the volume of water that the object carries. For this reason, it is very important that artificial intelligence can go beyond just performing the tasks assigned to it and reason abstractly.
IQ tests are used to measure abstract reasoning in humans, based on completing gaps between simple visual scenes. Although no explanation is given to individuals in these tests, participants can fill the gaps based on their daily experiences. In artificial intelligence, the application of such a test is not very functional as the yapay everyday experience yapay of artificial intelligence is very limited. For this reason, researchers have been inspired by the IQ tests to produce an abstract reasoning test that includes “progress” and ”color” and “size nitelik. In addition, in order to make artificial intelligence easier to understand the concepts in the tests, a set of artificial intelligence was prepared and artificial intelligence was trained.
The results obtained when the artificial intelligence tests were quite interesting. All of the artificial intelligence involved in the test had the error of applying the preparation set directly on the test. In contrast, some artificial intelligence had succeeded in completing the test with more than 75 percent accuracy. DeepMind explained that the most successful artificial intelligence progresses by clearly identifying the relationship between different images and eliminating potential answers through trial and error.
Taking the lessons from the mistakes After the first test, the artificial intelligence re-training, but this time the answers in the previous test is true or false, showing why the researchers, after this training began the second test process. According to the team’s report, when the correct explanation of the answers was made, artificial intelligence could reach 87 percent accuracy, but when the explanation was wrong, it remained at 32 percent.
As a result of this test, the researchers concluded that if the artificial intelligence generalizes (only tries to apply the data given to it), the abstract reasoning test will not be successful. The team, however, is hopeful of artificial intelligence that assesses potential answers in the first test. DeepMind is now working to improve the abstract reasoning ability of artificial intelligence by trying different training methods.