Software engineers at the University of Toronto were tackling a major challenge in artificial intelligence: training a neural network through written instructions instead of by showing it hundreds of examples. Then came the bigger finding: They realized the program had learned more effectively than they had taught it.
The researchers gave the program instructions for identifying visual images of hair – how to distinguish hair from the tree a person is standing in front of, for example, or how to tell hair from a hat.
They discovered that the neural net could perform 9 percent more accurately than it could during its lessons. In effect, the network taught itself how to improve on its own.
TRENDPOST: This incident marks a crucial step in achieving a key milestone in artificial intelligence – computers that learn by themselves.