By Peter Burrows It's a few days before the public launch of his new company, and Jeff Hawkins is excited but also concerned. "We don't want this to get overhyped," he says, slumping down with his head nearly on the conference table of his small Menlo Park (Calif.) office.
And worry about overexposure he should. For starters, Hawkins is a proven entrepreneur. The world's most famous designer of handheld computers, he co-founded Palm Computing and its offspring, Handspring (see BW Online, 10/21/04, "Wizard of the Wireless Future").
UNIFIED THEORY. This time, he's aiming at a far bigger opportunity than selling handheld gizmos. With his new company, Numenta, which he'll unveil on Mar. 24, he's trying to do no less than create machines that work just like the human brain. "This could radically change the way that computer systems work," says Harry Saal, a Silicon Valley veteran and one of a handful of investors in the company.
Certainly, Hawkins' pedigree, coupled with the vast implications of his new quest, will garner a lot of attention. But before you ask, the answer is no -- Hawkins' inquisitive brain hasn't taken him around the bend. Neuroscience, the study of the brain, has fascinated him since 1979, when he read a Scientific American article on the roots of intelligence. It started his obsession with finding the Holy Grail of neuroscience: a unified theory of how the brain works.
While neuroscientists over the years have parsed the problem into more digestible chunks -- say, how a neuron fires -- he says no one has put all the pieces together. "I tried several times to make this my career," says Hawkins, who three years ago created the self-funded Redwood Neuroscience Institute, where he spends most of his time. "Nobody had done the theoretical analysis of the brain."
"GENTLEMAN SCIENTIST?" Now, he thinks he has figured it all out and believes that he and an associate have come up with a way to translate his theories into electronic terms. Their company will license -- for free, at least for the first few years -- technology to others who may want to create what he predicts will be truly intelligent machines.
"He's moving from the neuroscience to the computer science," says Palm and Handspring co-founder Donna Dubinsky, who will serve as Numenta's chief executive. (Hawkins' title will simply be "founder.")
Of course, Hawkins' goal carries long odds, and critics say he has tried to take a shortcut by skirting peer review and letting the market, rather than more research, prove or disprove his theories. Some compare him to a certain 18th-century gentleman scientist -- a rich man using his personal wealth to pursue a radical vision. Of course, for every Ben Franklin, there were scores of others whose aspirations never panned out.
Not surprisingly, Hawkins' work is raising big questions from the hard-core science crowd. "He's a man of action, and he wants to get things done," says Terrence Sejnowski, head of the Computational Neurobiology Laboratory at the Salk Institute for Biological Studies in San Diego. "That's a different mode from most scientists, who are more interested in taking small steps and getting each step right."
NEURAL HIERARCHY. Hawkins' basic theory, laid out in his 2004 book, On Intelligence (see BW, 11/08/04, "Redefining Smart"), says the brain doesn't function as a miraculously fast processing unit, like a microchip. Instead, it represents a rather simple memory system that keeps track of all kinds of patterns over time -- whether they are sights, sounds, textures, or any other kind of input.
He believes this system is hierarchical -- that there are lower-level bits of brain that note specific details, which humans later synthesize into overall experiences. That allows someone, for example, to know that the combination of the touch of a finger on the shoulder and the sight of a masseuse's pillow means a massage will ensue.
By gathering an uncountable number of these patterns every instant of every day, the brain -- actually, the neocortex, or "the big wrinkly thing on top of the brain that does all the higher-level functions," as Hawkins puts it -- can predict what to make of any situation it encounters.
COMPUTERIZED CLAIRVOYANCE. If Hawkins' idea works, it could have an almost limitless number of applications. Attached to electronic eyes, such systems could pick out a terror suspect's face in a crowd as quickly as you could spot your mother -- a task today's computers can't handle.
The systems would also have military uses. Unmanned aircraft could go far beyond the rudimentary drones now in use. Or, the Army could outfit soldiers with a small device and various sensors enabling them to see, hear, and detect all manner of threats -- from an enemy creeping up from behind, to footsteps around the corner.
Plug in enough weather data, and one of Hawkins' computers might predict rain or snow -- not based on an actual forecast, but with a kind of electronic sixth sense akin to what experienced fishermen feel when they scan the skies.
FREE SAMPLES. Since 2003, Hawkins and co-founder Dileep George, who will serve as principal engineer at Numenta, have been working on software that incorporates such basic memory architecture. So far, they've developed a proof-of-concept program. By "showing" their program 90 simple line drawings -- one of a helicopter, another of a dog, for example -- they can teach it enough to identify another drawing of that object, even if it's poorly drawn or incomplete.
While science has already made similar efforts at machine vision, Hawkins insists this marks the first time a piece of software has had such predictive power. "If I see a cat behind a chair, I don't think it's half a cat," he says. And neither does his software, he claims.
While there is no undeniable proof of his sweeping unified theory, Hawkins has enough to take the next step: to develop the software and related tools so that engineers from various disciplines can try it out in their worlds.
INSTANT DEMAND. To encourage such widespread use, Numenta will initially charge nothing for its technology. Hawkins and Dubinsky, who have raised an undisclosed amount from a handful of friends and associates, will supply most of the $1 million-plus annual budget themselves.
They still need to figure out the licensing details. "The next step in the evolution of this is to create a profit motive for people to rally around," says Hawkins. Adds Gary Bradski, a machine-learning expert at Intel (INTC): "Even if he's wrong, his theory is better than nothing. And it's 'attackable' -- and that's a good thing."
Certainly, indications point to Hawkins receiving lots of interest. Raj Kent, a technologist with Lockheed Martin Advanced Technology Labs (LMT), says he has already spoken with Numenta about seeking grants for "cognitive computing" research the Defense Advanced Research Projects Agency will offer.
"THE MOST IMPORTANT THING." And Ajay Bakshi, a consultant in McKinsey & Co.'s health-care practice, read Hawkins' book and plans to contact him about using his ideas to help pharmaceutical companies discover new drugs. "Nobody was stepping back and looking at the big picture [of how the brain works]," he says. "Hawkins has given us a very big picture. It's a low-resolution picture -- but it will lead to a whole bunch of experiments."
That's just what Hawkins wants. "I don't need to run another company," he says. "But I think this is the most important thing I can do with my life. I'm trying to create a movement. Artificial intelligence had one for years, and neural networks had its time. But they were flawed theories. And I think we've got it right."
Experts caution against making such a claim too soon. But even if Hawkins finds only a small sliver of the Holy Grail he seeks, he'll add yet another industry-moving startup to his résumé. Burrows is Computer editor in BusinessWeek's Silicon Valley bureau