Room 3 · The Winters · 1989
Reading ZIP Codes
Reading ZIP Codes (Yann LeCun · AT&T Bell Labs, 1989): A convolutional network that recognizes digits.
In 1989 Yann LeCun and his team trained, with backpropagation, a convolutional network to read the handwritten digits of United States mail ZIP codes. It was an early version of the architecture that decades later powered image recognition, and it foreshadows what comes next.
In 1989 Yann LeCun and colleagues at AT&T Bell Laboratories published "Backpropagation Applied to Handwritten Zip Code Recognition" in the journal Neural Computation. The task was practical: read the handwritten digits of ZIP codes on United States mail, using images provided by the US Postal Service.
The network had a special design, called a convolutional network. Instead of connecting every pixel to every neuron, each neuron looked at a small patch of the image, and the same small set of weights was reused across all positions. In practice, this means the network learns a detector, for example for a short diagonal stroke, and applies it everywhere in the image. Later layers combine those simple detectors into more complex shapes. The design needs far fewer weights than a fully connected network and naturally handles a digit that is shifted a little to one side.
All the weights were learned with backpropagation, the method from earlier in this room, directly from the images. There were no hand-designed features, which was unusual at the time.
The idea grew into a family of networks known as LeNet. In the 1990s a system built on this work was used commercially to read the amounts on bank checks, and for years it processed a significant share of the checks written in the United States. In 1998 LeCun and colleagues published LeNet-5 and released the MNIST dataset of handwritten digits, which became a classic exercise for anyone learning machine learning.
Two decades later, the same core design, scaled up with much more data and graphics cards, won the 2012 ImageNet contest. That is the first piece of Room 4.
Ask your AI: how does a convolutional network find edges in an image? Ask it to explain with a small grid of numbers.
Source: doi.org/10.1162/neco.1989.1.4.541
See it in the museum →Cite this: Nazareno, I. (2026). Reading ZIP Codes (1989). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/reading-zip-codes-1989/