MNIST is the handwritten-digit dataset created by Yann LeCun and colleagues that contains 70,000 grayscale images.
The name MNIST stands for Modified National Institute of Standards and Technology database. It contains 60,000 training images and 10,000 test images of handwritten digits from 0 through 9. Each image is 28 by 28 pixels and is represented in grayscale.
MNIST became one of the most widely used introductory benchmarks for image-classification algorithms. Its fixed format and clearly defined labels made it easy to compare methods, teach neural-network concepts and test software implementations. Many early demonstrations of convolutional neural networks used MNIST or related digit-recognition tasks.
MNIST is sometimes mistaken for CIFAR-10, which contains color images in ten object categories, or Fashion-MNIST, a later replacement-style benchmark containing clothing images. MNIST is now considered relatively easy for modern systems, but it remains valuable as a transparent educational dataset and historical benchmark.