Right away, Richard F. Thompson knew Ted Berger was something special.It was 1972. Thompson was a Harvard professor in physiological psychology, at that time an area of particular strength within a powerhouse institution. Berger was a brand-new grad student, fresh from Union College in Schenectady, N.Y., where he’d graduated summa cum laude majoring in math and psychology and taken the prize for best scholastic record. Berger had come to Harvard to study the relationships between brain function and behavior. He didn’t lose any time validating Thompson’s first impression. Right off the bat, “Ted and another graduate student made a very spectacular discovery on the classical conditioning of the eye blink,” recalls Thompson. “Their paper was published in Science.”
Many academic scientists have long and productive careers without ever publishing a single article in Science, which, along with Nature, ranks as the most prestigious scientific journal in the world. By the time Berger received his Ph.D. in 1976, he had published an astonishing 10 papers and won the McKeen Cattell Award from the New York Academy of Sciences for his thesis research. (“Publishing is always the best sign of future success,” Thompson deadpans.)
Fast-forward a quarter century. Thompson and Berger are both professors at USC. Thompson is the Keck Professor of Psychology and Biological Sciences; Berger has joint appointments in biomedical engineering and biological science, and also directs the interdisciplinary Center for Neural Engineering.
Thompson still thinks his former pupil is something special. “Spectacular” is still the word he uses to describe Berger’s research. Indeed, it may be the single most spectacular research goal at USC today – or perhaps anywhere. Berger wants to open someone’s skull; implant a tiny, densely packed silicon computer chip; connect it to the brain; and let it take over cognitive function previously lost due to disease or injury.
“He wants to be the man who implants microchips between your ears. And the amazing thing is that he just might succeed,” a Wired magazine feature declared five years ago. Berger’s ambition to “create a bionic brain is bold, brash, and just a bit, well, mind-blowing,” the technology magazine opined. Berger’s project has come a long way since then.
This isn’t like a cochlear implant or an artificial retina or any other device stimulating inactive nerve fibers to resume functioning. No. This will be an artificial chunk of brain, something right out of a William Gibson cyberpunk thriller.
His science may be flamboyant, but Berger himself is anything but. At 52, his slicked-back hair is flecked with gray. Casual in polo shirt and slacks, he smiles a lot, talks easily and listens patiently. Sipping a glass of wine over lunch, he jokes about the neurons he’s killing with alcohol. “You couldn’t do this to a Pentium!” he says in praise of that wonderfully adaptive computer, the human brain.
Berger’s research is neither purely basic nor purely applied – but a rarely seen combination of both. It’s the perfect example of how academic research is supposed to advance scientific knowledge, educate students and benefit society. It has already sparked one successful commercial venture, and more spin-offs are forthcoming. Even if the quest to implant a brain chip should ultimately fail, Berger’s extensive contributions to the scientific understanding of how the brain, in particular the hippocampus, functions carries broad implications. His multiple-team methodology is a model of interdisciplinary research at its finest. The project has recruited an army of graduate students. Berger alone employs two postdocs and 14 graduate students – many drawn from the biomedical engineering core course he teaches. A half-dozen other faculty collaborators fill out the project’s ranks with their own troops of research assistants.
And it’s hard-edge research, raising questions for scientists and society. “We are on the brink of stretching the capabilities of the human race. I believe we will soon be able to connect the brain to computers or other devices,” Berger says. “We have to think about the implications.”
Berger’s vision has been compared to the movie Johnny Mnemonic, wherein a futuristic courier uploads digital data directly into his brain via an electronic port.
What Berger wants to do is actually harder. It involves more than just connecting circuitry to the brain; it calls for replacing damaged tissue with computer hardware that performs a function formerly carried out by neurons.
From left, USC neuroscientist Michel Baudry, computer scientist John Granacki and biomedical/electrical engineer Vasilis Marmarelis.
Photos by Michele A.H. Smith
Though Berger’s quest sounds fanciful, more and more people are becoming believers. Those at the National Science Foundation and the Defense Advanced Research Projects Agency are demonstrating their faith with large grants. Collaborators and observers alike exhibit growing enthusiasm.
“There’s no fundamental scientific reason that it couldn’t work,” says Thompson, of the bionic brain. “I get irritated with people who say we can never make computers as good as human [brains], because obviously we will.”
When Ted Berger was born, his father was studying electrical engineering at Purdue University. Dad went on to help pioneer transistor research at IBM; son Ted set high goals for himself too, hoping to make a difference in the world of neuroscience. “I started out wanting to understand everything about how the hippocampus works,” he says.
Berger’s interest has never really shifted from this region of the cortical brain found in all vertebrates. The cashew-shaped brain tissue plays a crucial role in learning and memory. Think of the hippocampus as a way station where experiences are initially processed, assessed and sorted. After a few days, those experiences deemed important will move on to long-term memory; the rest are destined for the brain’s dump heap. (Interestingly, when the hippocampus is removed – to treat epilepsy, for example – one loses the ability to form new long-term memories, but retains memories formed before the surgery.)
Especially intriguing to Berger was the hippocampus’ role in generating 3-D mental maps of one’s spatial position. Thus, a mouse with a damaged hippocampus can’t find its way around a maze. Neurologists believe the 20- to 50-percent loss in hippocampal volume associated with Alzheimer’s disease may explain why AD patients are prone to getting lost.
Berger seemed destined for a bright career in basic research, spending a couple of years at UC Irvine as a postdoctoral fellow, and another year as an Alfred P. Sloan Foundation fellow at the Salk Institute in La Jolla, Calif. In 1979, he accepted a faculty appointment in psychobiology at the University of Pittsburgh. As a rising star in Pitt’s top-notch neuroscience program, Berger joined the ranks of those intent on solving the puzzle of the so-called “black box.”
Since the 1950s, researchers the world over have painstakingly studied and documented the chemical reactions and associated electrical activity of the brain and its 100 billion nerve cells, or neurons.
Neurons communicate by sending electrical impulses to other neurons along networks of fibers, called axons; and neurons receive impulses through other long extensions, known as dendrites. Mapping this neural snarl has been likened to reaching into a mysterious black box, removing its contents piece by piece, and hoping that careful examination of each piece will reveal how the box works. Thompson calls this hope “naïve,” though it undoubtedly produces valuable basic research.
After 13 years at Pitt, Berger gradually came to a similar conclusion. “Eventually, I wanted to understand enough about neurons and the brain, and about networks of neurons, so that I could model them at a level that reproduces a cognitive function,” he says.
Berger began to ponder how to mimic what neurons did, even if he didn’t fully understand how they did it. A basketball player doesn’t need to be a rocket scientist to launch the ball on a perfect trajectory through the hoop, he reasoned. Then why should a neuroscientist need to understand every nuance of the brain before attempting a slam dunk?
“A neuron processes inputs into outputs, and I was focused on that,” Berger says.
He and his colleagues began bombarding live rat hippocampal neurons with a wide range of electrical impulses – all possible combinations – and recording the emerging electrical signals. Studying the rat hippocampus made sense: it’s essentially the same as a human one, but smaller. And hippocampal cells excised from rat brains retain much of their structure and can be kept alive with nutrients for a day or more.
The researchers traced how one neuron receives a sequence of digital-like pulses from another neuron ,how it transforms that signal into a new pattern of pulses and sends that along to a third neuron. Remarkably, Berger realized that “information to a neuron is embedded in the spatio-temporal pattern of input events.”
In other words, a neuron’s response to a given input depends on the timing of that input. Because a neuron usually receives inputs from multiple sources, the signal’s spatial direction also matters. Where one signal might not activate a neuron, a timed sequence of two signals will trigger that neuron to send its own signal down the line. As you increase the number of stimulations, the equation becomes increasingly complicated, and cataloging stimuli and responses becomes a daunting task.
Along comes biomedical/electrical engineer Vasilis Marmarelis. He and Berger began their collaboration at Pitt. The rat hippocampus research had piqued Berger’s curiosity about engineering in general, and non-linear systems modeling in particular. It was this evolving interest and the ties to Marmarelis and Thompson, both based at USC by then, that persuaded Berger to accept a faculty appointment here in 1992.
“USC is just so strong in engineering – especially engineering connected to the life sciences,” Berger says. “When you look at the neuroscience program here, half the people are engineers. That’s amazing. People just don’t realize how good USC is.”
Marmarelis’ specialty is non-linear systems modeling; he’s just about the best in the world at its application to biology.
But first, a few definitions: The relationship between two variables can be either linear or non-linear. Most often, it is non-linear – although we often mistake it for linear. For example, step on the gas and your car moves. A direct, simple and linear relationship, right?
It isn’t quite that simple, says Marmarelis. As your foot depresses the gas pedal, your car accelerates at an increasing rate. And there’s an almost imperceptible time gap between the depression of the pedal and the movement of the car. A teenager might try to shorten that slight hesitation by revving up the engine and popping the clutch. Clearly, the relationship between stepping on the gas and the forward motion of the car is a dynamic one: it changes with time. It is non-linear – like the relationships between various signals arriving to and departing from neurons in living slices of rat hippocampus.
Marmarelis has been creating mathematical equations to accurately represent the input and output activity of individual neurons and ever-larger groups of neurons. “You can take any two variables and examine the relationship and find a way to express it mathematically,” he explains. “A system is nothing more than a lot of different interacting variables.”
Based on Marmarelis’ mathematical neuron models, USC computer scientist John Granacki has built silicon chips that mimic neurons.
“When the chip receives real electrical signals as inputs, it processes them and sends out exactly the same signals that a real neuron would send,” says Granacki, who is director of the advanced systems division at the USC School of Engineering’s Information Sciences Institute. “It behaves just like a network of real neurons in the hippocampus.”
Granacki has fabricated circuits that take the place of about 100 neurons. To do anything useful in the brain, however, the researchers will need at least 10,000 neuron models on a chip. Berger and his colleagues designed such a chip and tested it in simulations. They’re convinced it will work; now they just have to build it.
But there’s a catch. The brain is often called the human body’s computer, yet it differs profoundly from man-made computers. Neurons send signals in milliseconds, or thousandths of a second. Today’s computer chips are about 100,000 times faster and getting faster all the time.
Computer chips are also serial. They do one thing at a time – very, very fast. The brain and nervous system, however, is parallel. Its signals traverse billions of channels simultaneously. In computer jargon, the brain is massively parallel.
Berger holds up a chip measuring about 1/8th of an inch, big enough to hold 100 neuron models. The two ends are studded with 25 pins each – a total of 50 pins.
“If you want to connect two chips together, you’ve got 50 pins here, and 50 pins over here,” he says. “You’ve got to connect every chip to every other chip. You can get 100 neuron models on the chip, but you need 10,0000 to do anything interesting in the brain. You’ve got a lot of wires, too many wires. I’m going to try putting something into your head that’s mostly wires and only a little bit of chip? I don’t think so.”
Enter Armand Tanguay. A professor of electrical engineering, he specializes in laser optics, optical devices and photonics. Instead of putting electrical signals on wires or pins, he replaces them with light signals.
It’s possible to place a microscopic laser on a chip; in fact, you can buy them off the shelf. Whenever the chip receives an input, it generates a tiny light signal. The signal passes through a lens to be received by an adjacent chip, making it possible to stack many chips together in a salami. Theoretically, a parallel-processing network of 10,000 interconnected neuron models would be no larger than a peanut.
“We know we can fabricate the chips with 10,000 neuron models inside them, and that it’s going to be small enough to put it inside your head,” Berger says. “That part’s easily within reach.”
The single biggest remaining hurdle is to figure out how to connect the silicon salami to living brain tissue – or “wetware,” as the engineers are fond of calling brains. The research team is well on its way to solving that problem, too.
Brain cells in the hippocampus are arranged in a double interlocking C-shape. Silicon chips, on the other hand, have a uniform geometry. So Tanguay has built a special interface chip that consists of an array of electrodes. The wires coming out of this interface chip conform to the peculiar geometry of hippocampus cells.
Still, there remains a fundamental problem. “How do you get neurons to live next to the hardware and communicate? You are putting electronics into a soup of ions,” explains Roberta Brinton, professor of molecular pharmacology in the USC School of Pharmacy. “Normally, you wouldn’t drop a radio or a Palm Pilot into some soup and expect it to keep working.”
Brinton is the key researcher working on the connection between neuron and silicon. Her solution is elegant. Except for the very tips of the electrodes, everything on the interface chip and on future implants will be insulated. The tips themselves are gold.
In culturing colonies of neurons on these electrodes, Brinton has found that the cells adhere much better to certain substances, such as aluminum or gold, than to others. Neurons, she discovered, are also naturally drawn to some materials. Brinton is exploiting this attraction to coax axons and dendrites from living neurons to grow to specific locations on chips.
In her lab, living neurons are sending impulses along axons that have twined themselves around gold electrodes sticking out of one side of Tanguay’s chip. The chip, encoded with Marmarelis’ mathematical neuron models, behaves just as a group of organic neurons would. It processes inputs and sends out impulses through electrodes on its opposite side, where living dendrites from another set of neurons have grown up like ivy. The researchers have tested the interface and are currently fine-tuning it.
“We can form contacts between microchips and functioning nerve cells for periods of days, but rarely for weeks. Years is out of the question right now,” says Berger. “Exactly how we’ll get the biological system to marry and interface with the non-biological one is one of the major hurdles left, but I’m confident we’ll succeed.”
photo by Michele A.H. Smith
Berger enlisted the help of USC chemist Mark Thompson to work on that problem. A specialist in nanochemistry, Thompson has an idea to use nanoscale DNA ladders, chemically cut from single strands of DNA, as scaffolding to hold dendrites and axons to the electrodes.
“We have a toolkit. We have an armamentarium of various strategies,” says Brinton. “If one approach doesn’t work, we’ll try another. But it is going to work.”
This isn’t merely Mouthing the Party line: it’s optimism grounded in real-world success. In 1999, Berger and biomedical engineer Jim-Shih Liaw (who has since left USC) produced a speech-recognition system based on the research team’s neuron models. It has proven better than human ears at picking out spoken words in a noisy environment. A review in Forbes magazine last year called the system an example of emerging “neurotechnology.” With the help of USC biomedical engineer Walter Yamada, the system is now under development for military and commercial applications with $3 million in funding. It isn’t based on a model of speech, says Berger. “It is based on a model of the brain.”
It’s no mystery why the neuron model works. When two people utter the same word, they don’t sound exactly alike, but their voices share common patterns. Neurons are extremely sensitive to patterns. “The brain is a superior computational device in many ways. One of those ways is pattern recognition,” says Berger.
The speech-recognition spin-off has led to two NSF grants totaling $2 million. If neural modeling can perform speech recognition so well, why not apply the concept to other information-processing problems? NSF administrators hope Berger and Granacki can develop the next generation of computer chip, able to identify faces, fingerprints or signatures substantially faster than current technology.
“Biologically inspired computing modules performing high-level pattern recognition will be a key aspect of future computing systems,” says Granacki. “We might be able to make a machine that can quickly scan medical images for signs of an abnormality.”
The Office of Naval Research’s interest in biologically based pattern recognition systems yielded another $2 million grant for Berger and Marmarelis to investigate brain mechanisms underlying sensor fusion – the ability to use multiple senses simultaneously to track events. For example, we often combine vision and hearing to determine who is talking in a noisy room. When necessary, we will trade off senses, looking away from the television to better focus on a phone conversation, or listening more intently if an obstacle suddenly blocks our view of a speaker’s face. While the brain handles sensor fusion functions readily, Berger says it has proven difficult to develop physical systems with this capability.
Meanwhile, DARPA gave Berger and his colleagues $4.7 million this year to support research on the brain-interface technology. This award was coupled with $3.5 million for new collaborators at Wake Forest University, who will attempt to implant Berger’s microchip model of the hippocampus into a primate brain within the next three years.
This work has brought Berger more than grants and productive scientific collaborations. He has formed close personal relationships with everyone on his team, but especially with Roberta Brinton. The two scientists were married in 1998.
These days Berger is more likely to be in a meeting, talking to the media or writing a proposal than hunched over a lab workbench. (He also lavishes attention on 12-year-old daughter Kimberly, who often tags along with him on campus. “Any spare time I have, I try to devote to her,” Berger says. “She’s just a really great kid.”) The scientific heavy lifting is being done by graduate students and postdoctoral fellows who will take the knowledge and technology they gather today and move it to higher levels tomorrow. Berger has been there, done that – and done it better than almost anyone else. What he seems to do best now is communicate.
“I can’t do all of this by myself,” he says. “Other people have to do their parts. How you go about that is a very interesting process.”
He has become a vocal advocate for interdisciplinary research. “I learned about photonics seven years ago. Now I know enough from working with people like Armand [Tanguay] to speak to them and listen to them. I watch their faces. I can tell when they understand, and when they don’t. I keep changing my words until they get that look.”
Marmarelis, Granacki, Brinton, Tanguay, Mark Thompson, Liaw and others like neuroscientist Michel Baudry and chemist Charles McKenna all have their own lines of inquiry. They’re some of USC’s most respected and successful scientists. All have a growing sense of excitement that Ted Berger’s science fantasy is headed for science reality.
“There is a grander vision, and to realize that grander vision requires a team of people to work together,” says Brinton. “Instead of each of us making bricks, we are all building the pyramid together. I expect to see the implant work. We will certainly see the application of this technology within our careers.”
It Takes a Campus
Building a bionic brain is not a single scientific problem but a series of challenges. USC and its Center for Neural Engineering use an interdisciplinary environment to work toward solutions.
Problem: How do hippocampal neurons work?
Solution: Ted Berger, with joint appointments in biomedical engineering and biological science, bombards live rat hippocampal neurons with electrical impulses and records the emerging electrical signals.
Next Problem: How do you model non-linear systems?
Solution: USC biomedical/electrical engineer Vasilis Marmarelis writes mathematical equations to represent the complex input and output activity of the neurons observed by Berger.
Next Problem: How do you turn those models into chips?
Solution: USC computer scientist John Granacki builds silicon chips that recreate the input and output activity modeled by Marmarelis.
Next Problem: How do you reproduce the brain’s massively parallel structure to create a small but workable implant?
Solution: USC electrical engineer Armand Tanguay replaces the wires on Granacki’s computer chips with laser optics to allow the connection of more than 10,000 neural models on a single network no larger than a peanut.
Next Problem: How do you get living neurons to interface with the implanted chip?
Solution: USC molecular pharmacologist Roberta Brinton discovers that neurons adhere better to gold than to other substances. She builds an insulated version of Tanguay’s chip with gold connecting electrodes.
Next Problem: How do you get the connections between the brain and the implant to last?
Solution: USC chemist Mark Thompson is using nanoscale DNA ladders, chemically cut from single strands of DNA, to hold the neurons’ dendrites and axons to the electrodes on Brinton’s insulated chip.