Deep Learning Revolution

Give a program 1.2 million labelled photographs, ask it to sort new ones into 1,000 categories of objects and animals, and allow it five guesses per photograph. How often does the right label miss all five? That share, the top-5 error, was the score in the 2012 ImageNet Large Scale Visual Recognition Challenge. The runner-up, which averaged several classifiers built on hand-designed image features, missed on 26.2% of the test photographs. A team from the University of Toronto, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, missed on 15.3%.