AT&T AI History: Support-Vector Networks and LeNet-5

AT&T laboratories contributed support-vector networks and LeNet-5 document recognition; the research spans Bell Labs and the later AT&T Labs.

kindCompany

Updated

Telecommunications and laboratory research

AT&T is a telecommunications company with a research history spanning Bell Labs and AT&T Labs. Its name covers a changing corporate structure: The historical Bell Labs research organization became part of Lucent in 1996, while SBC acquired AT&T Corp. and adopted the AT&T name in 2005. The 1990s research belongs to those historical laboratories rather than to a claim that today's corporation has always been the same legal entity.

Cortes and Vapnik's support-vector networks paper described a classifier that maps inputs into a higher-dimensional feature space and finds a separating decision surface. Its extension to nonseparable training data and handwritten-character experiments are concrete contributions beyond AT&T's role as a telecommunications operator. The journal identifies both authors with AT&T Bell Labs.

The LeNet-5 paper describes convolutional networks that learn character recognition from pixels, alongside a trainable system for reading bank checks. Three authors worked at AT&T Labs-Research; the paper documents the resulting system's commercial use in NCR check readers. This was research connected to an operating recognition application, rather than a later claim based only on a corporate affiliation.

Sources

  1. [1]
    Support-vector networks

    Machine Learning · Sep 1995

  2. [2]
    Gradient-Based Learning Applied to Document Recognition

    Proceedings of the IEEE · retrieved Oct 24, 2026

  3. [3]
    AT&T Labs | Our Legacy

    AT&T · retrieved Oct 24, 2026

  4. [4]
    The Historical Brands of AT&T

    AT&T · retrieved Oct 24, 2026

AT&T history

1998

■
Convolutional networks go to work reading checks

With Université de Montréal

LeCun's group had a convolutional network reading millions of bank checks in production, and shipped MNIST, the handwritten digit set a generation of researchers learned on. The architecture anticipated the convolutional networks that would win ImageNet fourteen years later, still waiting on the data and hardware that made that possible.

Proceedings of the IEEE
Research

1995

■
Support vector machines take the field

Cortes and Vapnik added a soft margin to the kernel classifiers Boser, Guyon and Vapnik had introduced three years earlier, letting the method handle data that could not be perfectly separated. For the next fifteen years, if you were doing machine learning, you were most likely doing some version of this.

Machine Learning
Research