MNIST and LeNet-5

The 46-page paper in the November 1998 Proceedings of the IEEE did two things that lasted. It described LeNet-5, a seven-layer convolutional network with two stages of convolution and pooling followed by fully connected layers, which is the template every later vision network refined rather than replaced. And it introduced MNIST, 60,000 training and 10,000 test images of handwritten digits, 28 pixels square, assembled from the US Census Bureau's employees and American high-school students.

MNIST was the field's common ground for fifteen years. Any new method could be tried on it in an afternoon, its error rate compared with the table in LeCun's paper, and its authors either encouraged or spared further effort. The best classical methods, support-vector machines among them, got below one percent error; convolutional networks got lower, which was the argument for them in the years when few were listening.

The paper also introduced graph transformer networks and end-to-end training of a whole document-reading pipeline, ideas that were ahead of the hardware. LeNet-5's descendants read the cheques; MNIST is still the first dataset every student trains on, and the digit instrument on this site runs a network of the same shape.