Scientists created AI with level of the four-year child
For many years writers, film directors and even scientists said to us that creation of really clever AI can bring by the end of all human race. And it is similar, we became closer to such reality. Though not so, to start to panic. Scientists from Illinoyssky university in Chicago and researchers of artificial intelligence from Hungary inspected AI which showed that level of its IQ we will compare with intelligence of the four-year child.
ConceptNet system with the open initial code, created within the MIT Common Sense Computing Initiative program, scientists checked Vexler’s test which defines an intelligence scale for preschool children and younger school students. The test allows to define strong and weaknesses of cognitive thinking, and when using computer AI as object of research checks understanding level the computer of language and allows to analyse process of its answers.
As reports a portal phys.org, results appeared not so optimistical as scientists counted on that. The test showed that level of computer AI is equivalent to level of the average four-year child, and it is much lower than that level which it is necessary to expect from children 5 – 7-year age.
The system ConceptNet represents a set of bases of the well-known concepts containing 275 000 concepts, describing the various relations between objects: «And is In», «And it is located in In» etc.
The main difficulties which were marked out at research, correspond with difficulties which can arise at the 4-year-old child. That the computer program understood tasks, scientists created the special linguistic modules translating text phrases in a set of keywords. For check of its mental abilities into it loaded the adapted questions from the third edition of test of Vexler. The answer was accepted in the form of the most probable communication between several concepts. For example, on a question «what color snow», the program should issue the terse answer – the name of this color.
The best results of ConceptNet showed in Dictionary blocks where it was required to explain a word meaning, and also in "Similarity" blocks where it was required to point to the general signs or communication between two taken objects.
The program showed the worst results in blocks where it was required to answer a question «Why?» or to "guess" on the basis of three directing helps, about what object there is a speech.