This is part of IEEE Spectrum’s SPECIAL REPORT: THE SINGULARITY
What do fruit-fly brains have in common with microchips? That’s not the setup for a bad joke; it’s David Adler’s life. Under Adler’s ultra sophisticated electron beam microscopes, advanced microprocessors with transistors far smaller than red blood cells have been reduced to their wiring diagrams. Now the noggin of the humble Drosophila melanogaster is next, as Adler is being courted by researchers at a neurobiology wing of the Howard Hughes Medical Institute to help them reverse engineer the human brain. They’re starting small, with the fruit fly.
Located in the green, rolling hills of Ashburn, in northern Virginia, the campus, known as Janelia Farm, has been described as a kind of Bell Labs for neuro-biology. Its task is solving what Adler calls the most important question in science: How exactly does the human brain do what it does? Lots of people are trying to answer this question, and there’s a growing impetus toward using high- definition brain scans to find out how the brain works.
“In a hundred years I’d like to know how human consciousness works,” says Janelia director Gerry Rubin. “The 10- or 20-year goal is to understand the fruit- fly brain.” It’s this difference between consciousness and brain that has neuro-science researchers stymied. The simplest system stores and processes information the same way the most complex system does; a primitive computer from 1986 works a lot like a supercomputer. Similarly, Rubin suspects that the human brain and the fruit-fly brain are separated only by degrees of complexity: “Just because it’s much more advanced doesn’t mean the basic wiring rules are different.” Right now, Janelia is working on a circuit diagram of the fruit-fly brain.
To that end, Rubin has stocked the Janelia campus with a collection of neuro-scientists, biologists, physicists, engineers, and computer scientists. The process resembles that of reverse engineering a microprocessor. It starts with a full-scale, three-dimensional wiring diagram of the fly’s brain, in which the density of neurons is substantially higher—“but not infinitely higher,” insists Adler—than the wiring in a high-end IC. “If we can get a circuit diagram of the human brain,” says Adler, “then we can understand what causes a lot of neurological disorders—depression, epilepsy, maybe even Alzheimer’s.”
Like an IC, the fruit-fly brain is subjected to logic and optical testing to derive its circuit diagram. With one approach, called neuronal electro physiology, researchers can record the electrical activity of neurons. “But the fly brain is even more complicated than an integrated circuit,” says engineer and group leader Eric Betzig. “With an IC, you know that every transistor fires the same—it’s either on or off. But the neurons in the brain don’t necessarily do that—they fire sometimes 20, sometimes 80, sometimes 100 percent.” So in addition to logic testing, the researchers also need to do imaging, and that’s where Adler and his amazing microscope come in.
A standard scanning electron microscope (SEM) images at about 10 million pixels per second. For comparison, a high- definition TV screen runs at 30 million pixels per second. In 2005, the Pentagon gave funding to California-based KLA-Tencor Corp., where Adler was then working, to invent a microscope that could operate at 1 billion pixels per second to verify circuit patterns on defense chips. Shortly after that, Janelia lured Harald Hess, a former colleague of Adler’s, to the campus to direct its applied physics and instrumentation group. When the Janelia team started looking into imaging, Hess called Adler. When he found out what Janelia was working on, Adler says, “it blew my mind.” Hess wasn’t interested in a microscope that could image at a paltry billion pixels per second. He wanted one that could process 10 billion per second.
To image the fruit-fly brain, the researchers use what they all refer to, gruesomely, as a “deli slicer”—the machine shaves 50-nanometer slices off the top of the infinitesimal fruit-fly brain “like slices of prosciutto,” says Betzig. (The same technique is used to reverse engineer microchips.) Then an electron microscope takes images of the brain slices, and these images are stacked carefully to form a 3-D virtual wiring diagram.
Slice, image, slice, image. Easy, right? Wrong.
Compared with an IC , even a tiny fruit-fly brain is a mess. One major bottleneck is the sample preparation: the brains must be sliced into perfectly even slivers before they’re imaged. Right now, that slice-and-image routine takes a whopping 10 months. The real time-waster isn’t the actual imaging—it’s the time it takes for each slice of brain to settle into place. Any movement, however slight, will make that hard-won image blurry. The fly brain is only about 300 micrometers on a side, but imaging one, even at 10 billion pixels per second, would take a whole day. You’re trying to image everything down to about 5 nm—about one-hundredth the size of what a regular lab microscope can resolve.
The storage requirements for the raw data alone are staggering: Adler estimates that scientists could rack up about a petabyte—that’s 1000 terabytes—of data for every day of imaging. Bear in mind that 1000 terabytes is for one fruit fly, with its sorry speck of a brain, and the biggest hard drive you can buy from a commercial vendor today holds only one terabyte of data. To get any good data, you’d have to compare hundreds of fruit-fly brains. Imaging hundreds of them at the speed and resolution of Adler’s technology would require a warehouse. “If nothing else,” he says, “you’re going to run out of space.”
Anyone over 30 remembers when a gigabyte of storage in one place was laughably sci-fi. It won’t be long before a 10-PB hard drive is as boring as today’s 100-GB hard drives. But this project doesn’t have as its goal merely collecting data; it is trying to establish the exact connections among the neurons and synapses of the tiny creature’s brain. And therein lies the big challenge. Each slice holds billions of pixels, and once every slice has been imaged, scientists have to piece them all back together to generate a 3-D wiring diagram. Adler compares the scale of the undertaking to trying to put together a real-time traffic map of North America from high-resolution satellite photos. “Now imagine that the United States is paved coast-to coast as densely as New York City,” he says. At the resolution necessary to see individual synapses, the data glut is crippling. “You have to turn that data glut into a wiring diagram that doesn’t take up 1000 hard drives,” says Adler.
He’s hoping that machine learning will compensate for the data glut. When American adventurer Steve Fossett disappeared in the Nevada desert last year, a virtual worldwide hunt ensued. People combed obsessively through Google Earth images for signs of the man and his plane. While telescopes and microscopes can image incredibly fine details, they still lack the all-important ability to interpret these images and throw away unnecessary data. Adler estimates that processing all the data from a 10-billion-pixel-per-second representation of the fruit-fly brain could take five years. So Mitya Chklovskii, Janelia’s resident theoretical neurobiologist, is trying to teach his computers to discriminate neurons from synapses, and synapses from axons. A computer that could store an enormous image for 5 minutes while it decides which data is relevant, Janelia director Rubin says, is a far more elegant solution than a bigger hard drive. “This would solve the data problem,” he says.
Let’s say all the engineering problems can be solved in the next five or 10 years. Could researchers then actually reverse engineer the human brain, creating its functional duplicate in silicon? Would consciousness and all its attendant joy, pain, insanity, and genius be freed from biological containment? Adler sees no reason why not. “The brain is the ultimate micromachine,” he insists. “The fact that it’s made out of meat is a red herring.”
His vision is a Google map of the human brain that incorporates not just Janelia’s circuit diagrams but also other work in neuroscience. Adler cites the work of Stanford neuroscientist Stephen Smith as “the first steps to finding the soul.” At Harvard’s Center for Brain Science, neuro-scientist Jeff Lichtman mapped mouse neurons by “painting” them with fluorescent proteins. Rubin believes he’ll live long enough to see an MRI-like device that measures function with such high-resolution output that neurons in fruit flies, mice, or even humans can be observed taking in and processing information in real time.
How would all these different systems work together to show us how the brain does what it does? With his 10- billion- pixel-per-second microscope, Adler is confident he’ll be able to produce brain-topography images like Google’s satellite views, resolving fine details in sharp focus. Smith’s cartography, on the other hand, he compares with Google’s map views, including street names. Rubin’s fMRI data would be like real-time traffic data. Layering these different maps atop each other, says Adler, could lead to a hybrid comparable to a Google map.
Such a Google-mapped brain, Adler says, could do more than let us understand and cure disease: it could lead to a map of human consciousness. And he believes that understanding the wiring of the brain could lead to transformative technologies. What are memories, he asks, but rewired patterns in our brains? “If you can understand how memories are formed,” he says, “you can create memories.” Just as today’s sophisticated circuit-editing tools can modify microchips after they’ve been manufactured and packaged, a brain-editing tool could perhaps one day modify the brain. Adler jokes about an application straight out of Total Recall: buying fond memories of a vacation instead of taking the actual trip.
In this heady context, the leap from reverse engineering the human brain to building a thinking machine doesn’t seem ridiculous. To Adler, the existence of human beings is proof enough that humans can be engineered. “When we study biology, we’re just studying a different version of nano technology—only it’s a more advanced nano technology.” But he quickly qualifies that statement: silicon is the wrong material, he adds. The nano technology we use today is static; we can move electrons around but not atoms, which means the chip doesn’t change when you use it. “We may not ever be able to get there using the silicon technology of moving electrons,” he says. “But someone could come along tomorrow and invent a different way of making a circuit that’s closer to what the brain does. Then, within 50 to 100 years, we’ll have something that can do what the brain does.”
But there’s nothing like a little healthy competition to speed up this timetable. Janelia isn’t the only player in the high-speed brain-imaging arena: both Harvard and the Max Planck Institute for Medical Research, in Heidelberg, Germany (where the 3-D SEM method of brain reconstruction was actually invented), are also working on the brain problem, and they compete heavily for milestones. The Harvard team may have solved the image-settling problem: they plan to adapt a conveyor-belt device used in the semi conductor industry as a continuously moving stage that allows an uninterrupted panoramic image, eliminating the need for time-wasting, steadying pauses.
Adler also consults for Harvard, helping its team push the limits of its existing SEMs by “supercharging them” to hit their full potential. Before a 10-billion-pixel-per-second microscope can be useful, he says, many other roadblocks have to be negotiated. So in the meantime, he takes these souped-up SEMs to the limits imposed on them by physics, not factory settings. That means, for instance, that a lab microscope with a default rate of 10 million pixels per second can jump to 100 million pixels per second after Adler is finished tweaking it.
Despite all the obstacles, the good news, Adler says, is that the fundamental physics of the superhigh-throughput electron microscope has been resolved. It’s no longer a science problem, he says: now it’s an engineering problem. Hess agrees. “Finding that one 65-nm shorted-wire defect in a Pentium chip and that one miswired neuron in a fruit-fly brain,” he says, are fundamentally similar problems. They’re counting on the inexorable climb of Moore’s curve to aid them in their process. Rubin describes the phenomenon in terms of his Ph.D. work sequencing a single yeast gene. “Thirtyish years later, DNA-sequencing machines are at the point where students are doing 100 of my Ph.D.s per second,” he says, laughing. “We’re at millisecond data acquisition. These are the kinds of advances we’ll need to make a map of the human mind.”
In Rubin’s mind, solving the fruit-fly brain is a 20-year problem. “After we solve this, I’d say we’re one-fifth of the way to understanding the human mind.”
For more articles, videos, and special features, go to The Singularity Special Report.