Artificial intelligence program that learns like a child
Artificial intelligence programs may already be capable of specialized tasks like flying planes, winning Jeopardy, and giving you a hard time in your favorite video games, but even the most advanced offerings are no smarter than a typical four-year-old child when it comes to broader insights and comprehension. It makes sense, then, that researchers at the University of Gothenburg have developed a program that imitates a child's cognitive development.
"We have developed a program that can learn, for example, basic arithmetic, logic, and grammar without any pre-existing knowledge," says Claes Strannegård. Starting from a set of simple and broad definitions meant to provide a cognitive model, this program gradually builds new knowledge based on previous knowledge. From that new knowledge it then draws new conclusions about rules and relations that govern the world, and it identifies new patterns to connect the insight to.
The process is similar to how children develop intelligence. A child can intuit, for example, that if 2 x 0 = 0 and 3 x 0 = 0 then 5 x 0 will also equal 0, or they could draw the conclusion that the next number in the series "2, 5, 8" will be 11. And the same kinds of intuition carry across to other areas, such as grammar, where it's easy to identify rules for standard verb conjugations from examples like sing becoming sang and run becoming ran in the past tense.
"We postulate that children learn everything based on experiences and that they are always looking for general patterns," Strannegård says.
The researchers' system, which they call O*, follows the principle of Occam's razor – that you should favor short and simple explanations over long and complex ones. It identifies patterns by itself and combines them with prior knowledge to solve problems.
Sometimes this will lead to errors, such as when children say "I brang my lunch" instead of "I brought my lunch," but O* managed not only to learn arithmetic from scratch, but also to perform above the average human level on propositional logic problems. And given enough information the researchers hope their program could learn and reason its way to correct conclusions across a range of domains without any need for a programmer to explicitly formulate which rules it should apply in a given situation.
"We are hoping that this type of program will eventually be useful in several different practical applications," says Strannegård. "I think a versatile household robot would be tremendously valuable, but we’re not there yet."
Strannegård and his colleagues presented a paper describing O* at the Seventh Conference on Artificial General Intelligence in August.
Source: University of Gothenburg