Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton described a deep convolutional network trained on ImageNet. Its results helped focus attention on the combination of data, compute, and learned representations.
Why it mattered: a system could learn useful visual patterns rather than relying entirely on hand-designed features.
The perspective: image classification identifies categories. It is one building block in the much broader problem of understanding the visual world. The original paper appeared in 2012.