From early 1996 to early 2002 I was head of the Image Processing Research Department at AT&T Labs-Research. Despite my group's name, much of our work and background was in machine learning, pattern recognition, and computer vision as much as in image processing.
We took this picture right before most members of the group left AT&T in early 2002.
back row (left to right): Vladimir Vapnik, Leon Bottou, Yann LeCun, Joern Ostermann, Hans-Peter Graf front row: Eric Cosatto, Tricia Green, Fu Jie Huang, Patrick Haffner.
As of 2006, Vladimir, Leon, Hans-Peter, and Eric are all at NEC Labs in Princeton, Patrick is at AT&T, Jie is a PhD student at NYU, and Joern is at Leibniz University in Hanover.
Former Members and Alumnis
My former group was the direct descendent of Larry Jackel's Adaptive Systems Research Department at Bell Labs. Created in 1985 the Adaptive Systems Research Department made important contributions to neural net hardware, learning theory, neural net algorithms, and pattern recognition, particularly handwriting recognition, and data mining. It was originally created as part of the Electronics Lab in Holmdel, NJ, lead by Chuck Shank, and later by Rich Howard.
The group produced not only the world's first neural net chip, but also the first (and may be the last) neural net chips to actually do something useful. Generations of chips included "Net32K" designed by Hans-Peter Graf, and the ANNA chip designed by Bernhard Boser and Edi Sackinger with help from Yann LeCun. Those chips used mixed digital/analog computation internally.
Produced in 1991, Net32K included 256 units with 128 weighted inputs each. The inputs and weights were binary, but units could be combined so as to allow 4-bit resolution on the inputs. Internal shift register allowed to perform convolutions with minimal off-chip communication. The peak speed was 320 billion multiply-adds per second (with binary vectors).
First tested in 1992, ANNA included 64 units with a total of 4096 weights with 6-bit precision and inputs with 4-bit precision. Optimized for convolutions, ANNA had a peak throughput of 4 billion multiply-add per second.
Both chips were integrated in various experimental boards and tested in PCs and Sun workstations.
(to be continued)
The following researchers were members of the group at various times: Bernhard Boser, Corinna Cortes, John Denker, Isabelle Guyon, Donnie Henderson, Larry Jackel, Edi Sackinger, Patrice Simard, Sara Solla,
Temporary members, postdocs, and visitors included among others Bernhard Schoelkopf, Alex Smola, Jason Stitson, Chris Watkins, Quen-Zhong Wu.
Yann LeCun, Professor
The Courant Institute of Mathematical Sciences
Room 1221, 715 Broadway, New York, NY 10012, USA
Copyright � 2000-2004 Yann LeCun.