The Google network that found cats
A network with a billion connections, trained on YouTube frames with no labels at all, produced a neuron of its own that responds to cat faces.
Why it matters
Unsupervised learning at sufficient scale generated meaningful concepts by itself, a result that reached every newspaper.
Training ran three days across a thousand machines with sixteen thousand cores. Besides cats the network produced face and human body detectors. The work showed that features need not be specified when data and computation suffice. The same year AlexNet proved the same point on a labelled set with far less hardware, and the cheaper route through graphics processors is the one that became the main line.