The success of deep learning in AI has been driven by the ability to train models on large datasets, such as ImageNet, which has led to significant improvements in image recognition tasks. In the past few years, the error rate in image recognition has dropped dramatically, with some teams achieving as low as 3% error, which is considered superhuman performance.