AlexNet (2012)

ImageNet Classification with Deep CNNs (Krizhevsky, Sutskever & Hinton — universally called AlexNet) entered the 2012 ImageNet competition and won by a margin so large it looked like an error: roughly a 15% error rate versus 26% for the runner-up.

What it showed: a deep convolutional network, trained end-to-end on two GPUs with ReLU activations and a then-huge labeled dataset, could demolish a decade of hand-engineered vision features. The ingredients weren't new — convnets dated to the 1980s — but data plus compute finally made the old idea work.