VeriFone uses Synaptics 'neural chip'

Electronic News, June 8, 1992 by Jonathan Cassell

SAN JOSE, CALIF.--Calling it the first application of neural network technology in a commercial product, Synaptics Inc. last week said its I-1000 Neural Eye Chip has been incorporated in a VeriFone Inc. system that identifies and "understands" check identification numbers.

"This is a simple application of the technology, but it's the first adapting neural technology to a real problem," said Carver A. Mead, chairman of Synaptics.

"The first microprocessor's first application was a calculator," Dr. Mead pointed out, adding "It's a simple thing, but that's the way all technology starts."

Dr. Mead, a noted semiconductor visionary, venture capitalist and teacher, co-founded Synaptics with Federico Faggin, a major influence in Intel's early microprocessor design work and a founder of Zilog.

Synaptics' I-1000 Neural Eye Chip mimics the functions of the synapses of the human brain. A neural net consists of thousands of analog processors that work in parallel and can recognize and learn patterns.

The I-1000 uses a heterogeneous architecture, combining an analog optical sensor and an analog neural net with a digital processor. The neural processor senses characters optically and sends the signals to its image classifier, an onchip network of 20,000 analog processors. The network finds patterns in the analog signals and determines their meaning. The recognized analog signals are then relayed to the I-1000's digital processing system that converts them to a digital format.

The I-1000 can read up to 1,000 characters per second and can perform up to 1 billion operations per second.

San Jose-based Synaptics, which is described as a maker of "tools for cost-effective intuitive computation," plans to manufacture the I-1000 in volume and sell it on the open market. Synaptics chose to design the I-1000 with older, 1.6-micron technology so it would be well within the technological capability of most fabs.

"The chip is highly manufacturable. It doesn't require any special processes," said Dr. Faggin, who serves as president and chief executive officer of Synaptics.

Dr. Faggin said Synaptics has already signed-on a Silicon Valley foundry to produce the I-1000, but declined to name the owner of the fab.

The first use of the I-1000 is in VeriFone's Gemstone Onyx check reader, a new type of check reader that actually sees and recognizes the identification numbers on checks. The Onyx's method of reading is different from conventional magnetic check readers that detect the flux on the magnetic character printed on the bottom of the check. In order to properly recognize the magnetic flux pattern, magnetic readers require the checks to be fed into the reader at a constant speed with a motorized feed that is subject to mechanical breakdown.

The Onyx, an 8.8-inch long, 3-inch wide black box with a check-reading slot on top, recognizes numbers regardless of scanning speed, allowing users to slide through checks themselves and eliminating all moving parts. In addition, the Onyx can recognize numbers of varying size and shape, enabling it to interpret the numbers on crumpled or folded checks. Finally, the Onyx costs considerably less than conventional check readers, whose use so far has been limited to clearing houses, according to VeriFone.

While Dr. Faggin describes this use of his neural net processor as "mundane," he said it was an important step toward popularizing the technology.

"If you can prove you can do something cheaper and better with technology, you remove the distrust in the user community," Dr. Faggin said. "Once you show it's not just for missiles, the new technology can start to take root in the world."

Dr. Faggin envisions that within two years I-1000-type processors will be used for purposes varying from dollar bill recognition in vending machines to product quality inspections on factory floors. Synaptics is currently studying a system for the U.S. Postal Service that can recognize addresses handwritten on envelopes.

Dr. Faggin said in 15 to 20 years a centralized telephone translation device that could convert one language to another in real time would be feasible. The same technology could later be miniaturized into a portable device, he added.

Even further into the future, Dr. Faggin said, neural net processors could act as an interface between humans and computers.

"You can envision computers that are absolutely friendly to users," Dr. Faggin said. "They can easily translate the fuzzy illogical world of humans into the logical world of computers and make them appear to be adaptable and forgiving.

"When you combine the two technologies, logical and intuitive, you will be able to make machines that you cannot conceive of today."

VeriFone's Gemstone Onyx will start shipping this month to supermarkets and large discount stores.