by Michio Kaku
Although these robot pets will have a large library of emotions and will form lasting attachments with children, they will not feel actual emotions.
REVERSE ENGINEER THE BRAIN
By midcentury, we should be able to complete the next milestone in the history of AI: reverse engineering the human brain. Scientists, frustrated that they have not been able to create a robot made of silicon and steel, are also trying the opposite approach: taking apart the brain, neuron by neuron—just like a mechanic might take apart a motor, screw by screw—and then running a simulation of these neurons on a huge computer. These scientists are systematically trying to simulate the firings of neurons in animals, starting with mice, cats, and going up the evolutionary scale of animals. This is a well-defined goal, and should be possible by midcentury.
MIT’s Fred Hapgood writes, “Discovering how the brain works—exactly how it works, the way we know how a motor works—would rewrite almost every text in the library.”
The first step in the process of reverse engineering the brain is to understand its basic structure. Even this simple task has been a long, painful process. Historically, the various parts of the brain were identified during autopsies, without a clue as to their function. This gradually began to change when scientists analyzed people with brain damage, and noticed that damage to certain parts of the brain corresponded to changes in behavior. Stroke victims and people suffering from brain injuries or diseases exhibited specific behavior changes, which could then be matched to injuries in specific parts of the brain.
The most spectacular example of this was in 1848 in Vermont, when a 3-foot, 8-inch-long metal rod was driven right through the skull of a railroad foreman named Phineas Gage. This history-making accident happened when dynamite accidentally exploded. The rod entered the side of his face, shattered his jaw, went through his brain, and passed out the top of his head. Miraculously, he survived this horrendous accident, although one or both of his frontal lobes were destroyed. The doctor who treated him at first could not believe that anyone could survive such an accident and still be alive. He was in a semiconscious state for several weeks, but later miraculously recovered. He even survived for twelve more years, taking odd jobs and traveling, dying in 1860. Doctors carefully preserved his skull and the rod, and they have been intensely studied ever since. Modern techniques, using CT scans, have reconstructed details of this extraordinary accident.
This event forever changed the prevailing opinions of the mind-body problem. Previously, it was believed even within scientific circles that the soul and the body were separate entities. People wrote knowingly about some “life force” that animated the body, independent of the brain. But widely circulated reports indicated that Gage’s personality underwent marked changes after the accident. Some accounts claim that Gage was a well-liked, outgoing man who became abusive and hostile after the accident. The impact of these reports reinforced the idea that specific parts of the brain controlled different behaviors, and hence the body and soul were inseparable.
In the 1930s, another breakthrough was made when neurologists like Wilder Penfield noticed that while performing brain surgery for epilepsy sufferers, when he touched parts of the brain with electrodes, certain parts of the patient’s body could be stimulated. Touching this or that part of the cortex could cause a hand or leg to move. In this way, he was able to construct a crude outline of which parts of the cortex controlled which parts of the body. As a result, one could redraw the human brain, listing which parts of the brain controlled which organ. The result was a homunculus, a rather bizarre picture of the human body mapped onto the surface of the brain, which looked like a strange little man, with huge fingertips, lips, and tongue, but a tiny body.
More recently, MRI scans have given us revealing pictures of the thinking brain, but they are incapable of tracing the specific neural pathways of thought, perhaps involving only a few thousand neurons. But a new field called optogenetics combines optics and genetics to unravel specific neural pathways in animals. By analogy, this can be compared to trying to create a road map. The results of the MRI scans would be akin to determining the large interstate highways and the large flow of traffic on them. But optogenetics might be able to actually determine individual roads and pathways. In principle, it even allows scientists the possibility of controlling animal behavior by stimulating these specific pathways.
This, in turn, generated several sensational media stories. The Drudge Report ran a lurid headline that screamed, “Scientists Create Remote-Controlled Flies.” The media conjured up visions of remote-controlled flies carrying out the dirty work of the Pentagon. On the Tonight Show, Jay Leno even talked about a remote-controlled fly that could fly into the mouth of President George W. Bush on command. Although comedians had a field day imagining bizarre scenarios of the Pentagon commanding hoards of insects with the push of a button, the reality is much more modest.
The fruit fly has roughly 150,000 neurons in the brain. Optogenetics allows scientists to light up certain neurons in the brains of fruit flies that correspond to certain behaviors. For example, when two specific neurons are activated, it can signal the fruit fly to escape. The fly then automatically extends its legs, spreads its wings, and takes off. Scientists were able to genetically breed a strain of fruit flies whose escape neurons fired every time a laser beam was turned on. If you shone a laser beam on these fruit flies, they took off each time.
The implications for determining the structure of the brain are important. Not only would we be able to slowly tease apart neural pathways for certain behaviors, but we also could use this information to help stroke victims and patients suffering from brain diseases and accidents.
Gero Miesenböck of Oxford University and his colleagues have been able to identify the neural mechanisms of animals in this way. They can study not only the pathways for the escape reflex in fruit flies but also the reflexes involved in smelling odors. They have studied the pathways governing food-seeking in roundworms. They have studied the neurons involved in decision making in mice. They found that while as few as two neurons were involved in triggering behaviors in fruit flies, almost 300 neurons were activated in mice for decision making.
The basic tools they have been using are genes that can control the production of certain dyes, as well as molecules that react to light. For example, there is a gene from jellyfish that can make green fluorescent protein. Also, there are a variety of molecules like rhodopsin that respond when light is shone upon them by allowing ions to pass through cell membranes. In this way, shining light on these organisms can trigger certain chemical reactions. Armed with these dyes and light-sensitive chemicals, these scientists have been able for the first time to tease apart neural circuits governing specific behaviors.
So although comedians like to poke fun at these scientists for trying to create Frankenstein fruit flies controlled by the push of a button, the reality is that scientists are, for the first time in history, tracing the specific neural pathways of the brain that control specific behaviors.
MODELING THE BRAIN
Optogenetics is a first, modest step. The next step is to actually model the entire brain, using the latest in technology. There are at least two ways to solve this colossal problem, which will take many decades of hard work. The first is by using supercomputers to simulate the behavior of billions of neurons, each one connected to thousands of other neurons. The other way is to actually locate every neuron in the brain.
The key to the first approach, simulating the brain, is simple: raw computer power. The bigger the computer, the better. Brute force, and inelegant theories, may be the key to cracking this gigantic problem. And the computer that might accomplish this herculean task is called Blue Gene, one of the most powerful computers on earth, built by IBM.
I had a chance to visit this monster computer when I toured the Lawrence Livermore National Laboratory in California, where they design hydrogen warheads for the Pentagon. It is America’s premier top-secret weapon
s laboratory, a sprawling, 790-acre complex in the middle of farm country, budgeted at $1.2 billion per year and employing 6,800 people. This is the heart of the U.S. nuclear weapons establishment. I had to pass through many layers of security to see it, since this is one of the most sensitive weapons laboratories on earth.
Finally, after passing a series of checkpoints, I gained entrance to the building housing IBM’s Blue Gene computer, which is capable of computing at the blinding speed of 500 trillion operations per second. Blue Gene is a remarkable sight. It is huge, occupying about a quarter acre, and consists of row after row of jet-black steel cabinets, each one about 8 feet tall and 15 feet long.
When I walked among these cabinets, it was quite an experience. Unlike Hollywood science fiction movies, where the computers have lots of blinking lights, spinning disks, and bolts of electricity crackling through the air, these cabinets are totally quiet, with only a few tiny lights blinking. You realize that the computer is performing trillions of complex calculations, but you hear nothing and see nothing as it works.
What I was interested in was the fact that Blue Gene was simulating the thinking process of a mouse brain, which has about 2 million neurons (compared to the 100 billion neurons that we have). Simulating the thinking process of a mouse brain is harder than you think, because each neuron is connected to many other neurons, making a dense web of neurons. But while I was walking among rack after rack of consoles making up Blue Gene, I could not help but be amazed that this astounding computer power could simulate only the brain of a mouse, and then only for a few seconds. (This does not mean that Blue Gene can simulate the behavior of a mouse. At present, scientists can barely simulate the behavior of a cockroach. Rather, this means that Blue Gene can simulate the firing of neurons found in a mouse, not its behavior.)
In fact, several groups have focused on simulating the brain of a mouse. One ambitious attempt is the Blue Brain Project of Henry Markram of the École Polytechnique Fédérale de Lausanne, in Switzerland. He began in 2005, when he was able to obtain a small version of Blue Gene, with only 16,000 processors, but within a year he was successful in modeling the rat’s neocortical column, part of the neocortex, which contains 10,000 neurons and 100 million connections. That was a landmark study, because it meant that it was biologically possible to completely analyze the structure of an important component of the brain, neuron for neuron. (The mouse brain consists of millions of these columns, repeated over and over again. Thus, by modeling one of these columns, one can begin to understand how the mouse brain works.)
In 2009, Markram said optimistically, “It is not impossible to build a human brain and we can do it in ten years. If we build it correctly, it should speak and have an intelligence and behave very much as a human does.” He cautions, however, that it would take a supercomputer 20,000 times more powerful than present supercomputers, with a memory storage 500 times the entire size of the current Internet, to achieve this.
So what is the roadblock preventing this colossal goal? To him, it’s simple: money.
Since the basic science is known, he feels that he can succeed by simply throwing money at the problem. He says, “It’s not a question of years, it’s one of dollars …. It’s a matter of if society wants this. If they want it in ten years, they’ll have it in ten years. If they want it in a thousand years, we can wait.”
But a rival group is also tackling this problem, assembling the greatest computational firepower in history. This group is using the most advanced version of Blue Gene, called Dawn, also based in Livermore. Dawn is truly an awesome sight, with 147,456 processors with 150,000 gigabytes of memory. It is roughly 100,000 times more powerful than the computer sitting on your desk. The group, led by Dharmendra Modha, has scored a number of successes. In 2006, it was able to simulate 40 percent of a mouse’s brain. In 2007, it could simulate 100 percent of a rat’s brain (which contains 55 million neurons, much more than the mouse brain).
And in 2009, the group broke yet another world record. It succeeded in simulating 1 percent of the human cerebral cortex, or roughly the cerebral cortex of a cat, containing 1.6 billion neurons with 9 trillion connections. However, the simulation was slow, about 1/600th the speed of the human brain. (If it simulated only a billion neurons, it went much faster, about 1/83rd the speed of the human brain.)
“This is a Hubble Telescope of the mind, a linear accelerator of the brain,” says Modha proudly, remarking on the mammoth scale of this achievement. Since the brain has 100 billion neurons, these scientists can now see the light at the end of the tunnel. They feel that a full simulation of the human brain is within sight. “This is not just possible, it’s inevitable. This will happen,” says Modha.
There are serious problems, however, with modeling the entire human brain, especially power and heat. The Dawn computer devours 1 million watts of power and generates so much heat it needs 6,675 tons of air-conditioning equipment, which blows 2.7 million cubic feet of chilled air every minute. To model the human brain, you would have to scale this up by a factor of 1,000.
This is a truly monumental task. The power consumption of this hypothetical supercomputer would be a billion watts, or the output of an entire nuclear power plant. You could light up an entire city with the energy consumed by this supercomputer. To cool it, you would need to divert an entire river and channel the water through the computer. And the computer itself would occupy many city blocks.
Amazingly, the human brain, by contrast, uses just 20 watts. The heat generated by the human brain is hardly noticeable, yet it easily outperforms our greatest supercomputer. Furthermore, the human brain is the most complex object that Mother Nature has produced in this section of the galaxy. Since we see no evidence of other intelligent life-forms in our solar system, this means that you have to go out to at least 24 trillion miles, the distance to the nearest star, and even beyond to find an object as complex as the one sitting inside your skull.
We might be able to reverse engineer the brain within ten years, but only if we had a massive Manhattan Project–style crash program and dumped billions of dollars into it. However, this is not very likely to happen any time soon, given the current economic climate. Crash programs like the Human Genome Project, which cost nearly $3 billion, were supported by the U.S. government because of their obvious health and scientific benefits. However, the benefits of reverse engineering the brain are less urgent, and hence will take much longer. More realistically, we will approach this goal in smaller steps, and it may take decades to fully accomplish this historic feat.
So computer simulating the brain may take us to midcentury. And even then, it will take many decades to sort through the mountains of data pouring in from this massive project and match it to the human brain. We will be drowning in data without the means to meaningfully sort out the noise.
TAKING APART THE BRAIN
But what about the second approach, identifying the precise location of every neuron in the brain?
This approach is also a herculean task, and may also take many decades of painful research. Instead of using supercomputers like Blue Gene, these scientists take the slice-and-dice approach, starting by dissecting the brain of a fruit fly into incredibly thin slices no more than 50 nm wide (about 150 atoms across). This produces millions of slices. Then a scanning electron microscope takes a photograph of each, with a speed and resolution approaching a billion pixels per second. The amount of data spewing from the electron microscope is staggering, about 1,000 trillion bytes of data, enough to fill a storage room just for a single fruit fly brain. Processing this data, by tediously reconstructing the 3-D wiring of every single neuron of the fly brain, would take about five years. To get a more accurate picture of the fly brain, you then have to slice many more fly brains.
Gerry Rubin of the Howard Hughes Medical Institute, one of the leaders in this field, thinks that altogether, a detailed map of the entire fruit fly brain will take twenty years. “After we solve this, I’d say we’re one-fifth of the way
to understanding the human mind,” he concludes. Rubin realizes the enormity of the task he faces. The human brain has 1 million times more neurons than the brain of a fruit fly. If it takes twenty years to identify every single neuron of the fly brain, then it will certainly take many decades beyond that to fully identify the neural architecture of the human brain. The cost of this project will also be enormous.
So workers in the field of reverse engineering the brain are frustrated. They see that their goal is tantalizingly close, but the lack of funding hinders their work. However, it seems reasonable to assume that sometime by midcentury, we will have both the computer power to simulate the human brain and also crude maps of the brain’s neural architecture. But it may well take until late in this century before we fully understand human thought or can create a machine that can duplicate the functions of the human brain.
For example, even if you are given the exact location of every gene inside an ant, it does not mean you know how an anthill is created. Similarly, just because scientists now know the roughly 25,000 genes that make up the human genome, it does not mean they know how the human body works. The Human Genome Project is like a dictionary with no definitions. Each of the genes of the human body is spelled out explicitly in this dictionary, but what each does is still largely a mystery. Each gene codes for a certain protein, but it is not known how most of these proteins function in the body.
