MNIST: The Dataset That Trained a Generation

In the late 1980s, the United States Postal Service had a problem. Thousands of mail sorting employees were reading handwritten zip codes on envelopes, a slow and error-prone process that cost the postal system millions of hours every year. Yann LeCun, then at AT&T Bell Labs, was handed a dataset of scanned handwritten digits from those envelopes, and was asked to teach a computer to read them.

LeCun's solution was a convolutional neural network, a system inspired by the visual cortex that would slide a small filter across an image, learning to detect edges, curves, and ultimately digits. His network, LeNet-5, achieved around 99% accuracy. But the real enduring gift came later: the cleaned-up benchmark dataset he created from that postal data.

In 1998, LeCun, Corinna Cortes, and Christopher Burges released the Modified National Institute of Standards and Technology dataset, MNIST. It contained 70,000 handwritten digit images, 28×28 pixels each, perfectly labelled. Simple enough to train on a laptop. Rich enough to separate good algorithms from great ones.

For over two decades, MNIST became the "Hello, World" of machine learning. Every PhD student, every researcher, every curious programmer training their first neural network started here. It became shorthand for a benchmark that was almost too easy"if you can't beat MNIST, you don't have a model." And yet it continued to expose real gaps: early methods struggled past 97% accuracy for years.

By the 2010s, MNIST was largely solved at 99.7% accuracy with modern ConvNets. But the dataset never retired. Today it still ships with nearly every ML framework as the default example. Somewhere right now, thousands of people are running their first neural network on LeCun's postal service dataset, unknowingly tracing a line from a New Jersey post office to the frontier of artificial intelligence.