Twirling slowly on a Cal Tech lab bench is a pinwheel that's the twinkle in Carver Mead's silicon eye.
Fresh from the fabrication line, the RET (as in retina) 30 is Carver's first real cut at mimicking the biology of the living eye on a sliver of silicon -- and it works.
The chip can see every curl, twist and rotation of the toy -- a feat that is beyond the vision systems of the most sophisticated of today's machines.
RET 30 is one of a handful of Mead-made chips designed to emulate living systems. Unlike the rigidly linear and binary structure of most computer chips, these microchips embrace the subtler, richer and infinitely more complex patterns of information processing that is used by living creatures.
"We're at a unique point in history," asserts Mead, a widely respected teacher and researcher in microchip circuitry design. "There was no medium for doing this sort of thing before."
Mead asserts that these "analog" nerve networks using VLSI (for Very Large Scale Integrated circuits) design will "do for neurobiology what gene splicing did for genetics -- do hypothesis and honest-to-god testing" that will allow microchips to see, hear, feel and "remember" in ways that transcend current designs.
These nascent chips could represent a new market potential in everything from visual pattern recognition to instantaneous speech recognition.
The industry "was using silicon as a cost-reduction for old ideas and they weren't using it as a medium for new ideas," says Mead. "It was just awful -- it was so dreadful to watch this opportunity pass us by. There are so many ways that are natural for silicon to do things for us."
This avant garde approach to "synthetic neurobiology" skips far beyond Silicon Valley's obsession with building smaller and faster versions of the same computer chips.
"A computer is such a trivial thing -- conceptually on the order of a few neurons," says Mead. "Lotus 1-2-3 represents just a tiny fraction of a neural level.
"Although its logic speed is a billionth of a second, the Cray supercomputer can't do what your eyes can do in a millisecond," says Mead. "That's a factor of 10 to the 10th. The only reason we don't build these systems is that we don't know how they work. But it's so powerful, we've got to understand what it is.
"Take 'neural wetware' as the metaphor. Evolution was able to generate systems within the constraints and properties of that wetware. You can view it as a highly engineered system because the life forms that weren't highly engineered were eaten up by the ones that were.
"Now I'm going to start with a technology I can implement and I'm going to evolve systems based on the constraints of that technology."
Not everyone is as upbeat about this approach as Mead.
"Carver's always enthusiastic," says Robert Noyce, vice chairman of Intel Corp. and coinventor of the integrated circuit, "but it will take him twice as long to do some of these things as he expects.
"You can't always say that the biological solution is the right one or else you'd have airplanes with wings that flap."
Still, Noyce says, "Carver is an absolutely brilliant man."
Mead, 52, is a compulsive teacher whose California Institute of Technology students -- known as 'Mead's Mafia' -- populate the top ranks of Silicon Valley's design labs. His effusive vocabulary of metaphors and analogies -- drawing on images from the quantum intricacies of semiconductor physics to the acoustical questions posed by echo location in bats -- is designed to impress the listener with the interdisciplinary nature of his work.
"Some people may think he's a nut," says venture capitalist Russel L. Carson of Welsh, Carson, Anderson & Stowe, "but he's dean of the semiconductor design industry. Anything he's interested in, we should be interested in."
"He can be as far as 10 years ahead of where anyone is, or wants him to be. But he's usually right," John Doerr, a venture capitalist with the firm of Kleiner, Perkins, Caufield, & Byers, said.
Mead's somewhat controversial reputation arises from his brand of academic evangelism and entrepreneurialism. While instrumental in offering insights into how best to exploit the potential of integrated circuits -- he goes the next step to popularize those insights and to encourage venture capitalists to build industries around them.
His fire-and-brimstone speaking style and guerrilla-like forays into new company start-ups grates on many in both industry and academe.
"He's brilliant, but I don't know sometimes whether he's explaining something or trying to sell me on it," gripes one Stanford researcher. "I think he likes preaching too much."
"He's a real hardball player in science and technology," says Lynn Conway, a collaborator with Mead and associate dean of engineering at the University of Michigan. "Mead's style is very much like an entrepreneur starting a company: going in spurts and then there's the crashing, hard-driving phase to drive it home. The product isn't a paper, though -- it's a new technology format."
Mead's complexities are reflected accurately in his medium of silicon, which he speaks of in the same way a sculptor might discuss clay or marble.
"A good design, a really good design, is beautiful, and people who do design can tell whether someone kludged it or whether they knew what they were doing," he says. "The silicon does have a way of telling you what it doesn't want to be. You can tell when you get it right."
Getting it right on silicon has become a task of staggering complexity. Advances in VLSI microchip fabrication enables designers to put more than a million transistors on a chip -- more than 10 times the number that could be placed on a chip less than five years ago.
Venture capitalist Doerr likens the VLSI design challenge to placing a flawlessly rendered map of the United States -- with every superhighway, street and back alley precisely located -- onto a piece of silicon the size of a thumbnail.
By decade's end, he projects, VLSI cartography will place the equivalent of a world map on a chip.
Mead's mission has been to give designers the concepts and the tools to cope with this complexity -- and to give industry the concepts and tools to create new designers.
In the mid-1970s, Mead devised the concept of a "silicon compiler" -- a computer-driven circuit "mapmaker" to aid designers and eventually make it possible for virtually anyone to create a custom-designed chip.
That led to Silicon Compilers Inc. in 1981 -- the first company to offer this technique of VLSI design and the forerunner of dozens of other efforts since.
Mead was a force behind the "silicon foundry" -- called "freedom of the press for silicon" -- by encouraging the semiconductor manufacturers to offer use of their facilities to freelance and academic chip designers to try out their designs -- turning the factories into silicon "printing presses."
He is perhaps best known for his collaboration with then Xerox researcher Lynn Conway for their 1980 book, "Introduction to VLSI Systems," which revolutionized the chip design industry.
Where chip design once had been a black art of physics, fabrication technology and a "connect the thousands of dots" mentality, the Mead-Conway work offered designers a structured methodology to manage the silicon real estate upon which their chips were built.
"The book was a watershed," says Albert Yu, assistant general manager of Intel Corp.'s microcomputer group -- and a former Mead student. "It irrevocably changed how people saw the design process. It changed the mindset of virtually every engineer in the Valley."
That work also gave Mead the framework to let him pursue the design of neural networks on silicon.
The first break came a few years ago, when on a "hunch" he thought he could use some special light-powered circuitry to do a simple model of a fly's eye. That failed -- but it launched a serious effort to use his expertise in VLSI design to build a silicon eye.
"If you're going to understand vision," Mead says, "look at biology because biology does it well."
Mead argues that current machine vision technology -- which relies on television cameras to give the computer "sight" -- is precisely the wrong way to approach the problem.
"A TV image is nothing like the real image," says Mead. "Just by starting with the wrong input device you've created a major problem for yourself."
For example, typical machine vision programs detect motion by comparing new images to previous ones. "That turns the problem into a correspondence problem," says Mead, because the computer has to use complex mathematics to figure out which new point corresponds with which old point.
"That's not how the eye works," he says. "The eye retains the image and continuously tracks the rate at which the image changes."
That's how RET 30 works -- which means that it can track motion more efficiently than any machine on earth.
"What is the problem with computers?" asks Mead rhetorically. "Computers can cope well with extremely well defined inputs and extremely well defined outputs. That's very different from what tasks face most of the world. Tasks are fuzzy. Computers have to be able to cope with the very fuzzy, very fluid undefined world out there."
Mead points out that most of our information processing is "unconscious" -- we aren't aware that we're seeing something: We just see it. Our minds then choose consciously to act upon that preprocessed information.
In essence, Mead has defined artificial intelligence not as how life "thinks" -- but how life perceives the world and organizes those perceptions for the mind.
"The reason artificial intelligence has failed so miserably as a discipline is because they've started in the middle," asserts Mead. " Nobel laureate and friend Richard Feynman once said Einstein was a giant because he had his head in the clouds and his feet on the ground -- but most of us aren't that tall so we have to choose.
"I've chosen to keep my feet on the ground.