To date AI learning is accomplished by processing incredibly large amounts of data.
In a research article published in Science, Lake and his team describe a model which learns “rich concepts” through a probability model built on Bayesian logic.
For example their program can “see” a wheel and from that “primitive concept” conceive of a two wheeled vehicle, a three or four wheeled vehicle, an engine powered vehicle, and so on. The program can in essence conceptualize possibilities from a basic example.
In the same way, the program can “see” a letter in the alphabet, identify the sub-parts and parts of that character and then create new characters. What is noteworthy about this ability is that the program-generated “new characters” are nearly indistinguishable from “new characters” generated by humans given the same task. The program arrived at the same end result as a human.
Alan Turing, the British mathematician who cracked the German Enigma code during the Second World War, predicted the creation of machines which could “think” as humans think. The test of whether or not a particular program or machine has this capability is called the “Turing test.”
When the results generated by the new AI program were compared to human efforts for the same task, the new program received an impressive 71% score on its Turing test.
It is literally unimaginable to conceive where this capability could take us.