History
LeCun, Cortes, and Burges compiled a very famous dataset with 60,000 training images of handwritten digits. What is the name of this dataset?
MNIST dataset.
Card 1 of 259. Answer: MNIST dataset.
MNIST dataset.
Card 1 of 259. Answer: MNIST dataset.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
MNIST, the Modified National Institute of Standards and Technology database. It holds 70,000 grayscale images of handwritten digits 0 to 9, each 28 × 28 pixels, split into 60,000 for training and 10,000 for testing. Yann LeCun, Corinna Cortes and Christopher Burges built it from NIST data in the 1990s.
NIST's Special Database 3 was written by Census Bureau employees and Special Database 1 by high school students, and NIST used them as training and test sets, making the test set much harder. MNIST mixed them, 30,000 training images from each, with different writers in training and test. Digits were scaled to fit a 20 × 20 box and centered in the 28 × 28 frame. LeNet-5, in LeCun's 1998 paper, reached 0.95% test error.
Modern convolutional networks get well below 0.5% error and even a linear classifier reaches about 90% accuracy, so MNIST is now a sanity check rather than a real benchmark. Fashion-MNIST, released by Zalando in 2017 in the same format, is a harder drop-in replacement.