A convolutional network on zip codes

A Bell Labs group trained a convolutional network to read handwritten postal codes and ran it on real United States mail.

Why it matters

The first back-propagation-trained network doing paid work, and the first convolutional architecture with learned features.

The constraints of the task were built into the architecture itself: local receptive fields, shared weights, subsampling. That cut the parameter count by an order of magnitude and let the network train on the data available. Training took three days on a SUN workstation; the same run takes minutes today. What the record does not claim. An earlier version said the system read about 10 percent of US cheques in the early 1990s; no source of the record says so, and the developers' own CVPR-97 paper dates the cheque reader's first deployment in a bank to June 1996, which is a record of its own.

Event record

Event date

1989

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

Learning theory

ID

evt-0159

Neural Computation volume 1, issue 4, 1989.

Sources

primary Backpropagation Applied to Handwritten Zip Code Recognition (publisher record)
MIT Press · Published 1989
archive Backpropagation Applied to Handwritten Zip Code Recognition
Neural Computation · Published 1989

Related events

Earlier

Builds on Back-propagation 1986
The network is trained by back-propagation.
- Backpropagation Applied to Handwritten Zip Code Recognition · Neural Computation
Builds on Neocognitron 1980
Convolution and weight sharing come from the neocognitron, but here they are trained by gradient.
- Backpropagation Applied to Handwritten Zip Code Recognition · Neural Computation

Later

Related ANNA: a convolutional network on an analog chip 1991links to this record
The chip runs a handwritten digit recognizer from the same Bell Labs group, a convolutional network previously trained and run on ordinary computers.
Related NIST's shared test on handwritten characters 1992links to this record
Among the participants were four AT&T systems, from the laboratory that trained the convolutional network on zip codes.
Related Falcon: a neural network against card fraud 1992links to this record
The second neural network to do paid work at national scale: there, mail and cheques; here, card payment authorisation.
Builds on A convolutional network reads cheques in banks 1996links to this record
The character recognizer is a convolutional network; the paper cites the 1989-1990 work on zip codes as its origin.
- Global training of document processing systems using graph transformer networks (L. Bottou, Y. Bengio, Y. Le Cun), CVPR-97, San Juan, 17-19 June 1997, pp. 489-494 · IEEE Computer Society; author copy on Léon Bottou's publications page
Extends LeNet-5 and MNIST 1998links to this record
The full account of the architecture shown on zip codes in 1989.
- Gradient-Based Learning Applied to Document Recognition · Proceedings of the IEEE
Related AlexNet at ImageNet 2012
AlexNet in 2012 is the same architecture, enlarged and moved onto a graphics processor.