CIFAR-10CIFAR-10

By the late 2000s, MNIST digits were too easy, but full photographic datasets were too big for the hardware of the day. At the University of Toronto, Alex Krizhevsky and Vinod Nair, working with Geoffrey Hinton, carved a middle path: 60,000 color images of just 32 by 32 pixels each, sorted into ten everyday classes such as airplane, cat, ship and truck.

The images came from MIT's 80 Million Tiny Images collection, which had been scraped from the web with noisy automatic labels. The Toronto team paid student labelers to verify every image by hand, and Krizhevsky documented the result in a 2009 technical report. The name honors CIFAR, the Canadian Institute for Advanced Research, whose funding had kept neural network research alive through its unfashionable years.

CIFAR-10 hit a sweet spot: hard enough that progress meant something, small enough to iterate on with a single GPU. Through the 2010s it served as the proving ground for one architectural idea after another, from dropout and batch normalization to residual networks and automated data augmentation.

Krizhevsky's experience squeezing performance out of CIFAR-10 on GPUs fed directly into AlexNet, the network that won ImageNet in 2012. In that sense the tiny dataset was a rehearsal space for the deep learning revolution.

Later chapters were instructive too. A 2018 replication study built a fresh test set and found every model scored several points lower, a caution about overfitting to a fixed benchmark. And in 2020 the parent 80 Million Tiny Images collection was withdrawn over offensive content in its automatic labels, while the hand-verified CIFAR-10 endured, still a standard first benchmark for new vision ideas.