Chapter 4 · The deep learning era (2006–2016)2012
ImageNet and AlexNet
AlexNet won ImageNet by a huge margin in 2012 and kicked off the deep learning boom.
Chapter 4 · The deep learning era (2006–2016)2012
AlexNet won ImageNet by a huge margin in 2012 and kicked off the deep learning boom.
Each bar is the best error rate in that year's ImageNet contest (rounded). For two years progress was slow. Then a deep network trained on GPUs cut the error by about ten points in one go.
ImageNet is a collection of over a million labelled photos in a thousand categories: dog breeds, tools, foods, vehicles. It gave researchers a shared test: name what is in the picture.
Name each drawing. Harder views make you slip too, which is what made ImageNet tricky. Your error rate is compared with the contest winners.
A teaching version: eight simple drawings and a small sample, so your score is only a rough guide. The contest used real photos, a thousand labels and top-5 scoring.
What happened in 2012?
What does a top-5 error mean?
Which two ingredients came together for AlexNet?