ImageNet: The Dataset That Triggered the Deep Learning Era

In 2006, Fei-Fei Li was a young professor at Princeton with an idea that seemed absurd: to build an image dataset so large that a machine learning model could learn visual concepts the way a child does, from vast exposure, not carefully engineered features. The ImageNet Large Scale Visual Recognition Challenge (ILSVRC) would eventually contain 14 million images across 1,000 categories. It would become the benchmark that broke AI open.

The bottleneck was labelling. ImageNet needed millions of images annotated correctly, a cat identified as a cat, a spatula as a spatula. Li's team turned to Amazon Mechanical Turk, the crowdsourcing marketplace where human workers complete small tasks for small payments. Over several years, workers labelled millions of images at roughly a cent per image. The entire dataset cost around $1 million to build, a fraction of what a traditional annotation effort would have cost.

The ImageNet challenge ran annually from 2010. Early winners achieved error rates around 25% using conventional computer vision techniques, hand-crafted features, SVMs, gradient methods. Then in 2012, a team from the University of Toronto, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, submitted a deep convolutional neural network they called AlexNet. Its error rate was 15.3%. Second place was at 26.2%. The gap was so large the organisers initially assumed there was an error.

AlexNet trained on two Nvidia GTX 580 GPUs for five days. It stacked seven layers deep, unprecedented at the time. Its use of ReLU activations, dropout regularisation, and data augmentation would become standard practice within a year. Krizhevsky's paper describing it has been cited over 90,000 times.

After 2012, every team competing in ImageNet used deep convolutional networks. Error rates fell each year: 11%, 6%, 3.5%, until models began outperforming humans. The challenge was retired in 2017 after object recognition was effectively solved. But the real prize was what ImageNet unlocked: proof that with enough data and compute, deep networks would learn features no human had engineered. The dataset built for a cent per label started a $100 billion industry.