|The Machine that Invents
From the ST Louis Post Dispatch
By Tina Hesman
Of the Post-Dispatch
Technically, Stephen Thaler has written more music than any composer in the world. He also invented the Oral-B CrossAction toothbrush and devices that search the Internet for messages from terrorists. He has discovered substances harder than diamonds, coined 1.5 million new English words, and trained robotic cockroaches. Technically.
Thaler, the president and chief executive of Imagination Engines Inc. in Maryland Heights, gets credit for all those things, but he's really just "the man behind the curtain," he says. The real inventor is a computer program called a Creativity Machine.
What Thaler has created is essentially "Thomas Edison in a box," said Rusty Miller, a government contractor at General Dynamics and one of Thaler's chief cheerleaders.
"His first patent was for a Device for the Autonomous Generation of Useful Information," the official name of the Creativity Machine, Miller said. "His second patent was for the Self-Training Neural Network Object. Patent Number Two was invented by Patent Number One. Think about that. Patent Number Two was invented by Patent Number One!"
Dining & Entertainment (72)
Automotive & Vehicles (61)
Home Improvement & Gardening (41)
Home Accessories (16)
Travel & Hotels (15)
Department Store (15)
...more on Ad Zone
Supporters say the technology is the best simulation of what goes on in human brains, and the first truly thinking machine.
Others say it is something far more sinister - the beginning of "Terminator" technology, in which self-aware machines could take over the world.
Thaler's technology was born from near-death experiences of dying computer programs. Its foundation is the discovery that great ideas are the result of noisy neurons and faulty memories.
The invention began to take shape in the 1980s. By day, the physicist worked at McDonnell Douglas Corp., where he wielded a powerful laser beam to crystallize diamonds. He built elegant computer simulations, called neural networks, to guide his experiments.
But at night, things were different. Shirley MacLaine and her ilk were all over the TV and on magazine covers talking about reincarnation and life after death and near-death experiences. It made Thaler wonder: "What would happen if I killed one of my neural networks?"
Neural networks can be either software programs or computers designed to model an object, process or set of data. Thaler reasoned that if a neural network were an accurate representation of a biological system, he could kill it and figure out what happens in the brain as it dies.
In biological brains, the information-carrying cells, called neurons, meet at junctions, called synapses. Brain chemicals, such as adrenaline and dopamine, flow across the junctions to stimulate or soothe the cells. In the computer world, there are switches instead of cells. The switches are connected by numbers or "weights."
So after work, Thaler went home and created the epitome of a killer application - a computer program he called the Grim Reaper. The reaper dismantles neural networks by changing its connection weights. It is the biological equivalent of killing neurons. Pick off enough neurons, and the result is death.
On Christmas Eve 1989, Thaler typed the lyrics to some of his favorite Christmas carols into a neural network. Once he'd taught the network the songs, he unleashed the Grim Reaper. As the reaper slashed away connections, the network's digital life began to flash before its eyes. The program randomly spit out perfectly remembered carols as the killer application severed the first connections. But as its wounds grew deeper, and the network faded toward black, it began to hallucinate.
The network wove its remaining strands of memory together, producing what someone else might interpret as damaged memories, but what Thaler recognized as new ideas. In its death spiral, the program dreamed up new carols, each created from shards of its shattered memories.
"Its last dying gasp was, 'All men go to good earth in one eternal silent night,'" Thaler said.
But it wasn't the eloquence of the network's last words that captured Thaler's imagination. What excited him was how noisy and creative the process of dying was. It gave Thaler ideas. What if, he asked, I don't cut the connections, but just perturb them a little?
Thaler built another neural network and trained it to recognize the structure of diamonds and some other super-hard materials. He also built a second network to monitor the first one's activities.
Then he tickled a few of the network's connections, and something began to happen. The tickling, akin to a shot of adrenaline or an electrical jolt in the brain, produced noise. In this sense, noise is not sound, but random activity. And the noise triggered changes in the network.
The result was new ideas. The computer dreamed up new ultra-hard materials. Some of the materials are known to humans, but Thaler didn't tell the network they existed. Other materials are entirely new, unknown to humans or computers before.
"A little elbow room"
When Rusty Miller went to lunch one day in 1998, he picked up a specialized computer magazine called PCAI journal. He flipped through the pages and came across a story about Thaler and his Creativity Machine inventing the ultra-hard substances. Instantly, Miller knew that Thaler had taken a step beyond other artificial intelligence technologies, such as fuzzy logic or genetic algorithms, he said.
The brilliance of Thaler's invention is the noise he introduces into the system, Miller said.
"Noise allows neurons to have a little elbow room to dream up new ideas," Miller said.
Other researchers have come to the same conclusion.
Good old-fashioned artificial intelligence uses human experts to input huge quantities of data and a list of rules to create a model, said Robert Kozma, a computer scientist at the University of Memphis. Kozma is experimenting with a similar technology.
The rigidity of traditional artificial intelligence technologies holds back creativity, Kozma said.
"This type of rule-based system is frozen. It's dead and cannot get to the essence of intelligence," Kozma said. "Creativity cannot be derived in a logical way, in a step-by-step fashion." You need a little noise to come up with good ideas, he said.
Human brains are also noisy places, said Dr. Walter J. Freeman, a neurobiologist at the University of California at Berkeley. A debate has raged for half a century about what the brain does with noise.
Many biologists see noise as just a nuisance or a necessary evil, Freeman said. The brain devotes many neurons to the same task so it can swamp out that random activity, those scientists argue.
But Freeman subscribes to an alternative theory - that noise is essential for the brain to function properly. Noise provides variability that allows organisms to adapt to new situations, he said.
Kozma has replaced the brain of a robotic toy dog with this new technology. The idea is to create a robot that can explore a new environment, such as the surface of another planet, without human guidance. NASA is funding Kozma's efforts.
Thaler believes that Kozma's research is derivative of his seminal work.
It's not merely noise that makes Thaler's Creativity Machines so ingenious, he argues. He has discovered a mathematical equivalent to the fleeting signals that work on neurons - a special kind of noise.
And Creativity Machines are their own best critics. In fact, they have critic networks built right in. The critics select the best ideas generated by the noisy networks and reward good work. The feedback helps the network dream up even better ideas.
Thaler, too, is engineering independent robots. A glossy, black, plastic cockroach named H3 could be the prototype for swarms of bunker-busting robots that could seek out, explore and use collective intelligence to defeat an enemy target. The U.S. Air Force has contracted Thaler to create such robots.
Robots, including Mars rovers, have been programmed with artificial intelligence before, Thaler said. But those robots require human engineers to program in leg movements and rules for getting around obstacles. Each unique encounter requires new programming, new rules, and time.
H3 gets no tutelage from Thaler at all. A sonar beacon beckons the robot, and H3's legs begin to flail. Every time the robot makes a movement that carries it closer to the signal, it learns the value of the move. Within a few seconds, the cockroach coordinates enough good moves to scuttle toward the signal.
But Thaler hasn't stopped with robots. Creativity Machines can solve just about any problem in any field, he says.
A Creativity Machine used two neural networks to study toothbrush design and performance. A brainstorming session between the two produced the idea to cross the bristles of the toothbrush for optimal cleaning. That toothbrush became the Oral-B CrossAction toothbrush.
In one weekend, a Creativity Machine learned a sampling of some of Thaler's favorite Top 10 hits from the past three decades and then wrote 11,000 new songs. Some are good, Thaler said. Miller confesses to being haunted by one of the melodies in a minor key. Other offerings are the musical equivalent of a painting of dogs playing poker, Thaler said.
But computer-composed music doesn't have to be bad. Human mentors with good taste could train a critic network to grade the Creativity Machine's songs, punish it for bad tunes and reward it for harmonious melodies. The feedback would hone the machine's composing skills.
Such a self-training system was the Creativity Machine's first invention, and the subject of Thaler's second patent.
Carmakers and security industries want to use machines to identify obstacles, pedestrians or intruders. Some machines can identify certain objects, but change lighting conditions or mist the lens with water, and the system falls apart.
Thaler spins a collection of toy cars, trucks and planes on an old turntable in his office while a Creativity Machine watches. The computer learns to distinguish Hummers from pickups and F-18s from 747s, no matter if the object is lit by a searchlight or sits in shadow or if rain spatters the windshield. The technology could alert drivers to whether they are about to back over a boy or a bicycle. Battlefield commanders might use similar technology to assess damage and decide whether to send in more bombs.
Machines trained to detect dangerous objects could replace humans at baggage screening stations or watch for suspicious behavior.
Thaler's first contract with the Air Force used a Creativity Machine to help design warheads that reconfigure the pattern of shrapnel scattering. That's important to limit collateral damage and to save money by tailoring bombs to destroy a target in one hit.
Thaler's machines engage in the guilty pleasure of reading supermarket tabloids. The networks learn how to write tabloid headlines. The "International Expirer" quickly became a hit on the Internet. But the computer reporters of the tabloid "have no shame," and generated such celebrity-skewering headlines that Thaler removed the Expirer to avoid libel and slander suits.
Spy agencies want to use Thaler's technology to map the Internet and detect unusual activity.
Thaler coined more than a million new English words by showing a network a list of words. It learned rules of spelling and pronunciation and generated new words. In one trial, the network came up with a name for one of Thaler's spinoff companies - Synaptrix. The words are nonsense now, but Thaler predicts that companies could use them to name products. The machine also liked "eggo." Too bad that one is already taken.
The technology is not ready for widespread commercial use yet, say some supporters.
"It's got extraordinary potential. Right now the holdup is packaging the technology as a tool that somebody can actually pull off the shelf and use," said Lloyd Reshard, the Weapons Platform Integration Team Lead at the Air Force Research Laboratory Munitions Directorate at Eglin Air Force Base. With other artificial intelligence technologies, "software is commercially available on the street, but if you want to apply a Creativity Machine to your problem, there's no software package you can go out and buy."
The Air Force is working with Thaler now to solve that problem, Reshard said.
"I might lose my job"
All of the possible applications for Creativity Machines make some people uneasy. The machines could easily supplant people for many mundane jobs, and Thaler predicts that some traditionally human-only jobs, including laboratory scientist, could be up for grabs. Computer chemists could soon design new compounds and figure out how to make them.
The machines could even be used to solve pressing societal problems, Thaler says.
The prospect is just too much for people who see machines as a possible threat to humans.
The normal human response is, "Don't want it. No thanks. I might lose my job," Miller said.
Or worse, sentient machines could decide that they don't need humans at all and do away with people. That fear is fueled by the plots of science-fiction movies, such as "The Terminator." In that movie, a satellite called Skynet became self-aware, saw humans as a threat and destroyed more than 3 billion people.
Sci-fi fans see similarity between Thaler's thinking machines and Skynet. There's even an eerie coincidence between the fictional satellite's Judgment Day - August 29, 1997 - and the date the patent for Creativity Machine became final - August 19, 1997.
But Thaler doesn't see the world ending at the hands of the machines.
"I can never imagine a world that looks like 'Terminator.' What do people want? Food. Land. Mates. Machines aren't interested in that," Thaler said.
Miller, who is in the business of protecting U.S. computers from foreign attackers, agrees that machines are not the real threat. He worries more about humans with malicious intent turning Creativity Machines into weapons. Other countries are already studying U.S. patents and experimenting with revolutionary technologies. Terrorists could follow suit, he says.
"If the U.S. doesn't wake up and pay attention, we're going to get smoked," Miller warns. "It's important for people to understand. It doesn't have anything to do with the business of business. It's about America."
Some people are threatened by the idea that machines could think like humans, Kozma said. They don't like the idea of computers out-creating humans, he said.
But Thaler's machines may never match the unique qualities of humans, no matter how clever they are at designing toothbrushes or warheads, Miller said.
Miller, a former ballet dancer and Green Beret, says he enjoys competing against Thaler's neural networks, even when they beat him. Miller will always have a toe up on the machines, he says.
"None of his computers can do ballet."
Reporter Tina Hesman