Deep learning · 2012 · Alex Krizhevsky, Ilya Sutskever & Geoffrey E. Hinton
ImageNet Classification with Deep Convolutional Neural Networks
Combine convolutional architecture, labeled data, GPU computation, ReLUs, augmentation, and regularization into a decisive empirical image-recognition result.
The central move
Combine convolutional architecture, labeled data, GPU computation, ReLUs, augmentation, and regularization into a decisive empirical image-recognition result.
Why it had to exist
Neural networks had useful theory and niche success, but large-scale vision required enough labeled examples, compute, optimization behavior, and regularization to make depth work together.
Where it leads
Backpropagation plus GPU scale → deep representation learning → modern vision and foundation models.