Neural nets win at images

In 2012 a neural network called AlexNet wins the ImageNet competition by a wide margin, cutting the error rate on labeling photos. Geoffrey Hinton's group, including Alex Krizhevsky and Ilya Sutskever, built it. Neural nets were old. What changed was a lot of labeled images, a graphics chip doing the math, and a few training tricks.

The contribution is learning from examples instead of rules written by hand. A vision system no longer needs a programmer to describe an edge. Show it enough pictures and it tunes its own numbers. Speech recognition and translation jump for the same reason in the same years.

This is software on chips that already existed. GPUs were built for games. The new use is matrix math. Without the 2007-era flood of photos and without the chip density from 1959 onward, AlexNet does not win.

2012 is a research result, not a product. The products arrive over the next decade.