Yann LeCun, with B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel of AT&T Bell Laboratories, published “Backpropagation Applied to Handwritten Zip Code Recognition” in Neural Computation, volume 1, issue 4, pages 541 to 551, in December 1989. The canonical DOI record and a verified scan of the original MIT Press article confirm the title, full author list, Bell Labs affiliation, journal, volume, pages, and date.
The paper showed that building constraints from the task directly into the architecture of a backpropagation network greatly improves its ability to generalize. The network was fed normalized images of digits rather than hand-engineered features, and it learned the entire recognition pipeline from raw pixels to final classification on real handwritten zip code digits supplied by the US Postal Service.
This is one of the earliest demonstrations of a convolutional neural network solving a real-world problem at scale, combining the architectural ideas of Fukushima’s Neocognitron with the backpropagation training of the 1986 Nature paper. The line of work led directly to LeNet-5 and, eventually, to the deep learning systems that now dominate computer vision.