Thought experiment: the plan to build a supercomputer replica of a human brain
Even by the standards of the TED conference, Henry Markram’s2009 TEDGlobal talk was a mind-bender.
He took the stage of the Oxford Playhouse, clad in the requisite dress shirt and blue jeans, and announced a plan that, if it panned out, would deliver a fully sentient hologram within a decade. He dedicated himself to wiping out all mental disorders and creating a self-aware artificial intelligence. And the South African-born neuroscientist said he would accomplish all this through an insanely ambitious attempt to build a human brain — from synapses to hemispheres — and simulate it on a supercomputer. Markram was proposing a project that has bedeviled AI researchers for decades, that most had presumed was impossible. He wanted to build a working mind from the ground up.
In the four years since Markram’s speech, he hasn’t backed off a nanometre. The self-assured scientist claims that the only thing preventing scientists from understanding the human brain in its entirety — from the molecular level to the mystery of consciousness — is a lack of ambition. If only neuroscience would follow his lead, he insists, his Human Brain Project could simulate the functions of all 86 billion neurons in the human brain and the 100 trillion connections that link them. And once that’s done, once you’ve built a plug-and-play brain, anything is possible. You could take it apart to figure out the causes of brain diseases. You could rig it to robotics and develop a new range of intelligent technologies. You could strap on a pair of virtual-reality glasses and experience a brain other than your own.
The way Markram sees it, technology has finally caught up with the dream of AI: computers are growing sophisticated enough to tackle the massive data problem that is the human brain. But not everyone is so optimistic. “There are too many things we don’t yet know,” says California Institute of Technology professor Christof Koch, chief scientific officer at one of neuroscience’s biggest data producers, the Allen Institute for Brain Science in Seattle. “The roundworm has exactly 302 neurons, and we still have no frigging idea how it works.” Yet over the past couple of decades, Markram’s sheer persistence has garnered the respect of people like Nobel Prize-winning neuroscientist Torsten Wiesel and Sun Microsystems cofounder Andy Bechtolsheim. He has impressed leading figures in biology, neuroscience and computing, who believe his initiative is important even if they consider some of his goals unrealistic.
Markram has earned all this support on the strength of his work at the Swiss Federal Institute of Technology in Lausanne, where he and a group of 15 postdocs have been taking a first stab at realising his grand vision — simulating the behaviour of a million-neuron portion of the rat neocortex. They’ve broken new ground on everything from expressing individual rat genes to the organising principles of the animal’s brain. And the team has not only published some of that data in peer-reviewed journals but also integrated it into a cohesive model so it can be simulated on a powerful IBM Blue Gene supercomputer.
The big question is whether these methods can scale. There’s no guarantee that Markram will be able to build out the rest of the rat brain, let alone the vastly more complex human brain. And if he can, nobody knows whether even the most faithful model will behave like a real brain — that if you build it, it will think. For all his bravado, Markram can’t answer that question. “But the only way you can find out is by building it,” he says, “and just building a brain is an incredible biological discovery process.” This is too big a job for just one lab, so Markram envisions an estimated 6,000 researchers around the world funnelling data into his model. His role will be that of prophet, the sort of futurist who presents worthy goals too speculative for most scientists to countenance and then backs them up with a master plan that makes the nearly impossible appear perfectly plausible. Neuroscientists can spend a whole career on a single cell or molecule. Markram will grant them the opportunity and encouragement to band together and pursue the big questions.
And now Markram has funding that’s almost as outsized as his ideas. On January 28, 2013, the European Commission awarded him €1bn (£844m). For decades, neuroscientists and computer scientists have been debating whether a computer brain could ever be endowed with the intelligence of a human. It’s not a hypothetical debate any more. Markram is building it. But will he be able to replicate consciousness? The EU has bet €1bn on it.
Ancient Egyptian surgeons believed that the brain was the “marrow of the skull” (in the graphic wording of a 3,500-year-old papyrus). About 1,500 years later, Greek philosopher Aristotle decreed that the brain was a radiator to cool the heart’s “heat and seething”. Although neuroscience has come a long way since then, the amount that we know about the brain is still minuscule compared to what we don’t know.
Over the past century, brain research has made tremendous strides, but it’s all atomised and highly specific — there’s still no unified theory that explains the whole. We know that the brain is electric, an intricately connected network, and that electrical signals are modulated by chemicals. In sufficient quantities, certain combinations of chemicals (called neurotransmitters) cause a neuron to fire an electrical signal down a long pathway called an axon. At the end of the axon is a synapse, a meeting point with another neuron. The electrical spike causes neurotransmitters to be released at the synapse, where they attach to receptors in the neighbouring neuron, altering its voltage by opening or closing ion channels. At the simplest level, comparisons to a computer are helpful. The synapses are roughly equivalent to the logic gates in a circuit, and axons are the wires. The combination of inputs determines an output. Memories are stored by altering the wiring. Behaviour is correlated with the pattern of firing.
Yet when scientists study these systems more closely, such reductionism looks nearly as rudimentary as the Egyptian notions about skull marrow. There are dozens of different neurotransmitters (dopamine and serotonin, to name two) plus as many neuro-receptors to receive them. There are more than 350 types of ion channel, the synaptic plumbing that determines whether a neuron will fire. At its most fine-grained, at the level of molecular biology, neuroscience attempts to describe and predict the effect of neurotransmitters one ion channel at a time. At the opposite end of the scale is functional magnetic resonance imaging, the favourite tool of behavioural neuroscience. Scans can roughly track which brain parts are active while watching football or having an orgasm, albeit only by monitoring blood flow through the grey matter: the brain again viewed as a radiator.
Two large efforts — the Allen Brain Atlas and the US National Institutes of Health-funded (NIH) Human Connectome Project — are working at levels in between these two extremes, attempting to get closer to that unified theory that explains the whole. The Allen Brain Atlas is mapping the correlation between specific genes and specific structures and regions in both human and mouse brains. The Human Connectome Project is using non-invasive imaging techniques that show where wires are bundled and how those bundles are connected in human brains.
To add to the brain-mapping mix, President Obama in April announced the launch of an initiative called Brain (commonly referred to as the Brain Activity Map), which he hopes Congress will make possible with a $3 billion (£2.2 billion) NIH budget. Unlike the static Human Connectome Project, the proposed Brain Activity Map would show circuits firing in real time. At present this is feasible, according to Brain Activity Map participant Ralph Greenspan, “in the little fruit fly Drosophila”.
Even scaled up to human dimensions, such a map would chart only a web of activity, leaving out much of what is known of brain function at a molecular and functional level. For Markram, the American plan is just grist for his billion-euro mill. “The Brain Activity Map and other projects are focused on generating more data,” he says. “The Human Brain Project is about data integration.” In other words, from his exalted perspective, the NIH and President Obama are just a bunch of postdocs ready to work for him.
Henry Markram has the tall build and tousled hair of a fashion model. Seated behind a clean desk in an office devoid of anything more personal than his white MacBook, he spends most of his days meeting with administrators, technicians and collaborators. The office is down the street from his wet lab and halfway across campus from the Blue Gene computer facility. Markram speaks of brain slices and microchips in detail, but he is not just a scientist in the conventional sense, stooped over a lab bench like Jonas Salk. He belongs to a new breed of telegenic research executives, a sort of J Craig Venter of the head. “I love experiments,” he says in a South African accent tweaked by more than a decade living and researching in Israel. “But I very quickly see that what I’m doing can be done far more efficiently.” Once the procedures for data collection are set, he believes, experiments can be outsourced or automated.
Understanding the brain writ large is what drives Markram. It has been his only serious interest since the age of 13, when his mother sent him from the Kalahari game farm where he’d spent his childhood to a boarding school outside Durban. In his first year there, he stumbled across some research on schizophrenia and other mental disorders and directed his youthful energy into studying the mind. “It was just amazing to me that you could have a little more or less of some chemical and your whole worldview would be different,” he recalls, smiling with boyish wonder. “If you can switch a chemical and your personality changes, who are you?”
To find out, he took up psychiatry at the University of Cape Town but swiftly grew impatient with the field. “I could see that this was not a science,” he says with a wave of his hand. “I didn’t see any future in it, grouping people by symptoms and prescribing whatever drug the pharmaceutical companies said.”
So he quit medicine and joined the only Cape Town lab doing experimental neuroscience, directed by a young researcher named Rodney Douglas. Even then — 1985 — Markram had formed his ambition to understand the whole brain. But he had to start at a much more granular level. Over a one-year period Markram performed nearly 1,000 experiments recording the effect of a neurotransmitter on neurons in the brain stem.
It was the beginning of his meteoric rise as an experimental neuroscientist. He got his PhD at the Weizmann Institute of Science, one of the leading research universities in Israel — “it was like landing in toyland,” he remarks with a broad smile — and went on to consecutive postdocs at the National Institutes of Health in Bethesda, Maryland, and the Max Planck Institute for Medical Research in Heidelberg, Germany. “My mantra is diversity,” he says, explaining his peripatetic years. “I clone my mentors. I copy everything they do, and then I innovate on top of it.” In 1995 he was recruited back to Weizmann as a senior scientist. In his new lab, Markram took up a technique that he’d learned from electrophysiologist Bert Sakmann at Max Planck, for which Sakmann and physicist Erwin Neher won the 1991 Nobel Prize in Medicine. The procedure called for a researcher to access a living neuron with a “patch clamp”, really just a micron-wide pipette, to directly monitor the neuron’s electrical activity. With his exceptionally steady hands, Markram was the first researcher to patch two connected neurons simultaneously, a feat that put him in a position to see how they interacted.
By sending electrical signals between neurons and measuring their electrical responses, he could test Hebb’s rule — neurons that fire together wire together, a fundamental neuroscience postulate. What Markram discovered was that the pattern of synaptic connections in a neural network is determined not only by whether neurons fire together but also by when they fire relative to one another. If an input spike of electrical current occurs before an output spike, the input connection is strengthened. If the input spike comes after the output spike, the connection weakens. In other words, Markram proved that the brain is attentive to cause and effect.
Markram published his groundbreaking results in a series of scientific papers, enough to earn him a full professorship by the age of 40. The lesson he drew from that success: he needed to set his sights much higher. “I realised that I could keep doing this for the rest of my career and I still wouldn’t really under- stand how the brain works,” Markram says. There were approximately 60,000 neuroscience papers published every year, only increasing the field’s fragmentation. What neuroscience needed, he decided, was an enormous collaboration, with research protocols co-ordinated so all the data would fire together — and naturally he thought he was the one to make it happen. His vision matched the ambition of one man who could fund it: neuroscientist Patrick Aebischer, the newly appointed president of the Swiss Federal Institute of Technology, tasked with making the campus a leader in computer science and biomedicine. In 2002 he recruited Markram, and in 2005 he bought him an IBM Blue Gene — one of the world’s fastest supercomputers.
From his position in Lausanne, Markram is doing four things simultaneously. He is running a wet lab that amasses data through experiments on brain tissue. Since 2005, he has been building a small-scale model and simulation of the rat neo-cortex (his initial Blue Brain project). He is now the co-ordinator of the lavishly funded Human Brain Project (HBP), spearheading a global initiative to co-ordinate data-gathering across labs worldwide. On top of all that, Markram is responsible for the simulation aspects of the HBP, building a virtual human brain from all the incoming data.
Markram’s Blue Gene supercomputer is a ten-minute walk from the Blue Brain wet lab, in a whitewashed room behind a sliding glass door. This is the second multi-million-pound supercomputer Switzerland has given him in ten years, with eight times more memory than his first. There are four racks of processors, each in a metal locker about the size of a washing machine. The loud drone of air-conditioning serves as a constant reminder that computing has a lot to learn about efficiency from the 20-Watt human brain.
The Blue Gene will simulate Markram’s brain model — the model that uses the experimental results he has collected over ten years of industrial-strength science at Lausanne, as well as the studies he did at Weizmann. But the model isn’t just a massive database. Markram understood that it would take trillions of dollars, not billions, to model every part of the human brain. “Other people in the field were saying that we didn’t know enough to start,” he says. (The Allen Brain Atlas’s Christof Koch, for one. Markram’s first mentor, Rodney Douglas, for another.) “What I realised was that you can get to the unknowns indirectly. It’s like putting together a puzzle with missing pieces. If you can see the pattern, you can fill in the gaps.” Markram calls the process predictive reverse-engineering. He claims that it has already allowed him to anticipate data that would have taken years to generate in a wet lab. For example, only about 20 of the 2,970 synaptic pathways in one small part of the rat neocortex have been experimentally measured. Detecting a pattern, he was able to fill in parameters for the remaining 2,950 pathways and to observe them working together in a simulation. Then he measured several in the wet lab to validate his reverse-engineered data. The simulation proved correct.
Markram is a man seemingly mired in contradiction. He wants to know mankind by studying the rat. He wants to industrialise experimentation and one day make lab work obsolete. He insists on exhaustive biological detail yet strives to make the most general models possible. But if you listen carefully — filtering out his relentless boasting — the apparent contradictions resolve into complementary strategies: without a dependable experimental base — focused on one species to which researchers have unlimited laboratory access — detailed modelling wouldn’t be possible. And without modelling and simulation, all that knowledge about the brain would amount to an incoherent storehouse of trivia. But with a multilevel model of the rat brain as a template, scientists might find a rule governing how neurons connect and chart only a few, on the basis of which they could fill in the remainder. “A unifying model is a powerful accelerator, since it helps you prioritise experiments,” he says. “I’m very pragmatic. The question is, what’s the minimum I need to know about the brain to reconstruct all of it?”
Markram continues to battle a chorus of nay-sayers. The eminent neuroscientist Moshe Abeles of Bar-Ilan University in Israel points out that the brain “differs from one individual to another, and in some respect also differs in each of us from day to day. Our ability to understand all the details of even one brain is practically zero. Therefore, the claim that accumulating more and more data will lead to understanding how the brain works is hopeless.”
Abeles didn’t keep his opinion to himself while Markram’s proposal was under review as one of six finalists for the billion-euro European Flagship Initiative grant. In the Israeli newspaper Haaretz last year, he proclaimed: “It is obvious the researchers won’t be able to keep their promise. It’s robbing the public purse on one hand and sabotaging the future of science on the other.”
Criticism also came from Rodney Douglas, who moved to Lausanne’s arch-rival, ETH Zurich, in 1995. “We need variance in neuroscience,” he declared at a session of the Swiss Academy of Sciences in January 2012, spreading alarm that with a billion euros Markram could achieve a monopoly on the field.
"Rodney Douglas’s resistance is a farce," Markram responds, sounding more sad than angry. "It’s envy, it’s ego. He’s at the end of his career, measuring a piece of a circuit, and he still doesn’t know what it’s doing." As if to prove Markram’s point, Douglas — who declined to be interviewed — will retire in July.
Koch believes envy is responsible for most criticism of Markram. “This is not a zero-sum game,” he says. “It isn’t that Henry is going to get a billion euros or neuroscience is going to get it. The money comes out of the European infrastructure. If it doesn’t go to his modelling facility, it might bail out a Greek or Italian bank.” Koch is sceptical of Markram’s ten-year time frame, but that didn’t keep him from spending three days this spring in Lausanne, co-ordinating their respective research programmes. “I like his vision,” Koch says. “The guy has cojones.” The distinguished University of Manchester computer engineer Steve Furber, inventor of the ARM processor, is even more fully won over. “There aren’t any aspects of Henry’s vision I find problematic,” he asserts. “Except perhaps his ambition, which is at the same time both terrifying and necessary.”
Markham thinks the greatest potential achievement of his sim would be to determine the causes of the approximately 600 known brain disorders. “It’s not about understanding one disease,” he says. “It’s about understanding a complex system that can go wrong in 600 different ways.” Rather than uncovering treatments for individual symptoms, he wants to induce diseases in silico by building explicitly damaged models, then find workarounds for the damage. Researchers have done the same with lab animals for decades, observing behaviour after giving them lesions. The power of Markram’s approach is that the lesioning could be carried out endlessly in a supercomputer model and studied at any scale.
And the view wouldn’t just be from the outside. Neuroscientists could see the flow of neurotransmitters and ions while experiencing the delusions. “You want to step inside the brain,” Markram says. He’ll achieve this by connecting his model to sensor-laden robotics and recording what the robot is sensing and “thinking” as it explores physical environments, correlating audiovisual signals with simulated brain activity as the machine learns about the world. A neuroscientist could then play back those perceptions as distorted by a damaged brain simulation. In an immersive 3D environment, a researcher could see the world as a schizophrenic while watching what is going on in the schizophrenic’s mind.
In hype-driven contexts ( such as his 2009 TED talk), Markram has hinted at the possibility that a sim embodied in a robot might become conscious. Hardwired with Markram’s model and given sufficient experience of the world, the machine could actually start thinking (à la Skynet and HAL 9000). Although that has gained him a following among science-fiction enthusiasts, he separates such speculations from the hard work of doing real science. When pressed, he shows a rare touch of modesty. “A simulation is not the real thing,” he says. “I mean, it’s a set of mathematical equations that are being executed to recreate a particular phenomenon.” Markram’s job, simply put, is to get those equations right.
He plans to give the European Union an early working prototype of this system within just 18 months — and vows to “open up this new telescope to the scientific community” within two and a half years — though he estimates that he’ll need a supercomputer 100,000 times faster than the one he’s currently got. Ever the optimist, he believes Moore’s law (and the EU) will deliver him that raw power in about ten years’ time. However, he’ll also need far more data than even his industrial-strength Blue Brain lab can collect.
Shortly after arriving at Lausanne, Markram developed workflows that extracted experimental results from journals, strip-mining thousands of neuroscience papers only to find the data was too inconsistent to use in a model. For a while, that looked like one of his biggest hurdles. But he’s since been building standardised protocols for many of the labs participating in the Human Brain Project. His timing may be just right, with the data glut expected from the Allen Brain Atlas, the Human Connectome Project and the Brain Activity Map. According to Brown University neuroscientist John Donoghue, a key figure in the Obama-sanctioned initiative, “the two projects are perfect complements. The Human Brain Project provides a means to test ideas that would emerge from Brain Activity Map data, and Brain Activity Map data would inform the models simulated in the Human Brain Project.”
One of the few people with experience simulating the human brain, University of Toronto psychologist Randy McIntosh is also tentatively optimistic about Markram’s project. “I think it’s possible,” he says. “I think of the Human Brain Project in the same way one should have considered the Human Genome Project, where the thought was that once the genome was sequenced, we would solve genetic-based disease and understand the genetic basis of behaviour. We’re nowhere near that, but in moving towards that goal, a huge number of insights and innovations came.”
Genomics has proven that biology, like astronomy and physics, thrives on big data. In the 21st century, going big is the way of all science. The brain is due for a billion-euro enlargement.
Image2: The Human Brain Project will eventually need an astronomical amount of memory and computational speed — at least 100 petabytes of RAM and an exaflop respectively.