Creating machines that truly mimic the mind has long been the Holy Grail of computer scientists. And like the path to that mythical prize, the quest for artificial “brains” that can balance a checkbook or recognize flaws in aircraft engines with all the aplomb of a human is strewn with failed attempts But now there is a promising new contender. The approach goes by the name artificial neural network, because it works by duplicating the neural structure of the brain and it is already proving its mettle in perhaps the sternest test of all—the cutthroat competition for investment capital.
A new wave of entrepreneurs anxious to turn the fledgling science into profitable products is on the scene, and dozens of small startup companies are springing up across the country (see chart page 7). Already it is possible to buy neural net programs that predict the movement of the Standard & Poor’s 500 or diagnose heart disease.
This technology may eventually become as ubiquitous as the transistor or the telphone,” proph esies Nobel Prize winner Leon Cooper, founder of neural network startup Nestor Inc. Adds Jasper Lupo, director of the Defense Advanced Research Projects Agency’s Tactical Technology Office, “I believe that this technology which we are about to embark upon is more important than the atom bomb.”
In fact, the field has gotten so hot in the financial community that some scientists are worried about people getting burned. “Promises are being made today by commercial neural network companies that could throw the industry into another Dark Ages,” warns Stanford University professor Bernard Widrow, who has been flirting with the field of neural networks for three decades.
But right now companies are not listening to Cassandras like Widrow. Instead, entrepreneurs are focusing on the technology’s potential to take over many sophisticated “thinking” tasks now performed by people. Rather than simply performing calculations or operations in sequence as do today’s computers, neural nets can make thousands of choices simultaneously. Moreover, unlike existing expert systems, which need to be programmed with a set of rules before they can solve problems, a neural network system can learn by trial and error. Given a set of prototype problems, it writes its own set of decision-making rules, and thus alleviates the need to be programmed. The result is a key leap forward, a machine that can learn.
The most beguiling—and, many scientists believe, most promising—form of neural network is one which relies not on conventional chips with linear circuits, but contains thousands of interconnected elements, that copy the pattern of neurons and synapses in the human brain. Called synthetic neural networks, these systems might even be able to perform such complex tasks as maintaining defense systems like the Strategic Defense Initiative. “Only about 10% of the data we now gather can be processed in real time [by the advanced artificial intelligence systems currently in use],” states DARPA’s Lupo. But DARPA hopes that neural networks currently under study at the agency will have the capability to almost instantly digest the other 90% of the incoming data, allowing quick target identification and equally swift reactions.
Construction of a synthetic neural network, however, is a daunting and long-range goal—one that is too far removed for all but the most stout entrepreneurs and venture capitalists. But there is a shortcut to building a working neural network: using software on existing digital computers. And it is this shortcut that dozens of entrepreneurs and small companies have taken, creating products that could only have been dreamed about 15 years ago, and in the process, bursting open the coffers of venture capitalists.
‘This fledgling industry will take our company to the $1 billion-a-year level by 1992,” predicts a hopeful Robert Hecht-Ni, founder of two-year-old Hecht-Ni Neurocomputer Corp. And Ed Rosenfeld, market analyst and editor of the Intelligence Newsletter in New York, states that the two best-funded neural network startups alone—Hecht-Ni Neurocomputer and Synaptics— have garnered almost $30 million in private venture capital in just two years.
Leading this group of new companies are neural network pioneers like Nestor and Hecht-Ni. Nestor’s Leon Cooper, for example, got started in neural networks back in 1972 (the same year he won the Nobel Prize for his work in superconductivity) when a graduate student at Brown University challenged him, saying that much had been discovered about neurons, but no one had any idea where memory was. Cooper recalls that his scientific feathers were ruffled by such “incompetence” among his biologist colleagues—and he spent the following decade designing a theoretical model of learning and memory with Charles Elbaum, then chairman of the physics department at Brown.
In order to harvest useful products from their theory, Cooper and Elbaum founded Nestor Inc. in 1980. Today the company is acknowledged as a leader in the field of software for artificial neural networks. Nestor has several products on the market, which it will either customize to a specialized application, or sell directly to a client wishing to do the application development in-house. The approach seems to work; Nestor’s impressive list of customers includes General Electric, Ford. Motor, Chemical Bank, Hughes Aircraft, Lockheed, Morgan Stanley, and Salomon Bros. And although the startup is still operating at a loss— living off of the capital brought in by the company’s initial public offering in December 1983, Cooper expects to realize his first profit early in 1989.
Nestor’s secret, contends Cooper, has been to not shoot too high, but to concentrate on bite-sized, do-able problems, and to represent its capabilities fairly. For instance, the company’s first product was a handwriting recognition system. Although traditional artificial intelligence systems can—and have—been programmed to recognize handwriting, neural networks perform the task better, faster, and cheaper, Cooper says. The neural network teaches itself the patterns and minor nuances of handwritten words, rather than recognizing only the patterns that have specifically been programmed into it, as a traditional computer would. The Nestor Writer has been converted for use by several banks, among them Chemical Bank, to read the handwritten amounts on checks.
In another project, a collaboration with Owen Carroll at the State University of New York, Nestor has developed a system that can identify whether a person’s heartbeat is abnormal. “We are not claiming that our system can diagnose heart problems as well as a doctor,” Carroll explains. But it can “identify when a doctor should be called in to diagnose a heart problem.” And it costs about $40 per diagnosis, compared to a typical fee of about $150 for a doctor’s examination.
Neural net purists might scoff at Nestor’s software-only approach. But other startups have taken the more difficult plunge into actual neural network hardware, albeit modestly, producing hardware accelerators that boost the capacity of digital computers. The leader in this area is Hecht-Ni Neurocomputer Corp. (HNC), founded by mathematician Robert Hecht-Ni.
The seed for HNC was sown in 1968 when Hecht-Ni read a paper by Stephen Grossberg, now the director of the Center for Adaptive Systems at Boston University. “I was inspired by his insights into how neural networks could be modeled mathematically,” Hecht-Ni recalls. Hecht-Ni set out to translate Grossberg’s descriptive equations into computer programs and shortly had a primitive software simulation of a neural net- work running on a minicomputer.
But Hecht-Ni’s simulation ran too slowly for his liking. “I quickly learned that specialized neural hardware [that could accelerate processing on a digital computer would be needed to do significant work,” Hecht-Ni explains. “At TRW we had a Neural Network Interest Group, sort of a Saturday morning research project,” he recalls. “We predicted that in five years everyone would want this technology if only the necessary tools were available to ease the development effort.”
But TRW had just liquidated its consumer electronics division under which neural-based tools would have fallen. So in 1986, Hecht-Neilsen and TRW engineer Todd Gutschow struck out on their own, own, and, with the financial backing of Simon Ramo (the “R” in TRW) among others, founded HNC.
Within a year, HNC had developed a set of accelerators for neural-based systems—today, these sell for around $10,000 per set. Although HNC is still privately held and will not reveal income figures, market analyst Tom Schwartz of The Schwartz Associates, Mountain View, Calif., predicts that HNC will sell just under $2 million worth of these development systems in 1988. And Hecht-Ni claims that HNC has “well over 50% of the market” and is growing at a rate that will take it to the billion-dollar mark by 1992.
The success of pioneers Nestor and HNC has not gone unnoticed—and now it seems like every scientist with a good neural net idea is taking advantage of venture capitalists’ infatuation with this hot new field. In fact, there have been so many startups in the past year (at least seven) that one new entrepreneur thought that the best way to set his company apart from the pack was to name it Yet Another Neural Network Computer Company (YANNCC, pronounced “Yank”). The firm was founded early in 1988, after general partner Michael Viehman heard Robert Hecht-Ni give a lecture about his system. “I could immediately see that there was a better way to build [hardware accelerators] that was faster, but used less expensive components,” Viebman explains. And today he is hard at work trying to prove that assertion to companies designing their own neural-based applications.
Most of the new entrepreneurial companies need the promise of profit within a few years in order to stay afloat But some have convinced their financial backers that the real payoff will come from a longer-view approach to the field. Rather than marketing software emulations for today’s niche markets, these startups are focusing on designing hardware for actual synthetic neural networks—which many contend are the only hope for tackling larger, more difficult problems such as endowing a computer with the ability to recognize novel visual or auditory stimuli.
One of the most promising startups to take this tack, Synaptics Inc., is driven by chief scientist and chairman Carver Mead. Mead, a professor at the California Institute of Technology and already famous for coinventing the silicon compiler (software that automates the construction of digital microchips), joined Gary Lynch, a neurobiologist, and microprocessor inventor Federico Faggin in founding Synaptics in August 1986.
Synaptics hopes to actually duplicate the mechanics of human vision to make a computer that can “see.” The first task in doing so is to map out the circuitry between the eye and the brain—a process Mead estimates will take 10 years. Accordingly, he does long-range theoretical work at CalTech while advising Synaptics on shorter-range product plans. So far, Synaptics has produced neural-based microchip sensors modeled on the eye and ear. But these are only experimental—and the company has no current schedule for releasing real products before 1990.
Surprisingly, such a long-term outlook is heartily endorsed by Avalon Ventures, the major financier behind Synaptics. “All those companies out there doing software simulations of neural networks are in the educational toy business,?’ says Kevin Kinsella, managing partner at Avalon, who seldom looks for his company’s investments to mature in less than seven to 10 years. “To do truly useful things you have to have hardware.” Even as money pours into these new companies, however, some scientists worry that the whole field is being oversold, that venture capitalists and the public alike are being led down the garden path with hype and grand promises. Perhaps the most outspoken critic is artificial intelligence pioneer and MIT professor Marvin Minsky. While Minsky admits that neural-based technology is “undoubtedly the way of the future,” he claims that irresponsible entrepreneurs are misleading people by overinflating their expectations. “There is nothing wrong with this field other than some of the people in it,” he alleges. Even successful entrepreneur Cooper objects to the hyperbole of some of his competitors—the claim, for example, that neural nets will be as smart as Hal in 2001. He insists that neural technologies must first solve bite-sized problems and that those who claim otherwise are charlatans. am a firm believer in neural networks, but when I hear some of the things that people say about them, I hold onto my wallet. What we must strive for now are practical systems that do useful things.”
Oregon-based freelancer R. Colin Johnson is author of Cognizers— Neural Networks and Machines That Think (John Wiley & Sons).