A machine capable of thinking for itself and expressing emotion is being developed in Switzerland
The group, which included the philosopher Huw Price, the cosmologist Martin Rees and the founder of Skype, Jaan Tallinn, were setting up a research centre, the Cambridge Project for Existential Risk, to work out the answer, and to study those potential one-off species-ending events that are the stuff of scientists’ nightmares.
To whet the public’s appetite for destruction, they drew up a shortlist of man-made apocalyptic scenarios, which included climate change, biotechnology and nuclear war.
But it was the final item on the list, artificial intelligence (AI), that caught the imagination. 'What happens if computers reach and exceed human capacities to write computer programs?’ Price and Tallinn asked. 'The moment that machines are able to develop even more intelligent machines would result in an “intelligence explosion”.’ (The man who first realised this, Jack Good, who worked with Alan Turing at Bletchley Park, suggested that the creation of a machine of such sophistication would be 'our last invention’, as ever-smarter robots left humanity far behind.)
The headlines were dramatic: killer robots? cambridge brains to assess ai risk. It was all wonderful publicity – unless, of course, you happened to be leading a scientific project of unprecedented ambition, aiming to develop supercomputers of hitherto unseen power to model the entire human brain in all its intelligent, emotional complexity.
Henry Markram, 50, is in charge of the Human Brain Project (HBP) at the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland. By creating a computer simulation of an entire human brain, Markram’s team aims to discover 'profound insights into what makes us human, develop new treatments for brain diseases and build revolutionary new computing technologies.’
As the Cambridge Project for Existential Risk was fretting about the future, Markram was waiting to hear whether the HBP had won €1 billion in EU funding – a sum that would transform it over 10 years into one of the biggest scientific quests in the world.
The race for that funding has been run since 2009, with 23 proposals vying to become the first two EU Future and Emerging Technologies (FET) flagship projects. The flagships idea was set up to foster research into radical scientific concepts across the continent. On January 28, at a conference in Brussels, Markram’s project was announced as one of the winners (the other was a pan-European project to develop the wonder material graphene). The next day I meet Markram in Lausanne, where he explains the aims, origins and potential implications of the project.
'Simulating the human brain is the key tool we want to use to reconstruct the brain systematically, to piece the pieces together, derive the biological rules, test them,’ he says. 'Now, of course, when you do that, and you capture the detail and can account for everything, the thing starts to behave like a piece of real tissue – and we can see that even at a small scale. If we connect it to a robot, for example, we anticipate that the robot will learn with its brain to develop different behaviours, higher brain functions, become cognitively capable, potentially intelligent.’
For Markram, that is a prospect that should inspire wonder, not fear. From his point of view, achieving a perfect reconstruction of the brain will signal that we have reached a profound understanding of how it works, and of how to repair it when it goes wrong. That, above all things, is the mission to which he has devoted his career. 'I was always interested in curing the brain,’ he says in English that remains tinged by a South African accent even after decades in the international melting-pot that is top-level academia. But his early career – delving into our consciousness through psychiatry – proved frustrating. 'So I decided I needed to go into the brain to find out how it works.’
Markram studied electrophysiology at Cape Town University and started by examining electrical processes in living animals, then focused in, working at ever-tinier scales, down to individual cells, and from there to the genes that carry the information to build those cells – 'the whole trajectory’, as he calls it. It was only when he had completed that trajectory that he realised that he had undertaken it in the wrong direction, because he wanted to understand how the brain worked as a whole. So he reversed his course, scaling his ambitions up, not down, and he began a journey to 'link the causal chain from genes to proteins to cells to connections to brain regions to cognitive behaviour to cognition’.
The HBP very nearly didn’t happen. Markram was a brilliant young electrophysiologist on his way to MIT when, in 2002, days before he flew to America, he got a call from Patrick Aebischer, a neuroscientist who had recently become head of EPFL. Aebischer had big plans for EPFL, then a solid if unspectacular institution framed by Lake Geneva below and the snow-capped peaks of the Alps above. He wanted to attract a man of ambition and big ideas, a man who might be able to pull EPFL into the big league. So he called Markram and made him an offer he couldn’t refuse.
'At MIT I had a big [funding] offer at the beginning,’ Markram says. 'But I realised it would be a massive uphill battle at the end. I came here. They said, “You want to do this? We’re behind you.”’
Three years later, in 2005, EPFL bought an IBM supercomputer, the Blue Gene L. It has two processors, like those that feature in home computers, on each chip, and two chips on each so-called 'compute card’, 16 of which slot into a 'node card’. There are 32 node cards in each 'rack’, a slate-grey cabinet that stands about 7ft high, and whose drab colour is enlivened only by the brightly coloured cables harnessing all the humming processing power. It didn’t come cheap. Each rack cost about $2 million. In partnership with IBM, EPFL acquired four.
The Blue Gene L allowed Markram to launch the Blue Brain Project (BBP). The basic premise was simple: to simulate the cortical column of a rat. Cortical columns are the building blocks of the brain. They consist of thousands of individual brain cells (neurons), which are composed of nuclei surrounded by nerve fibres (axons), and branched, tree-like structures (dendrites). Synapses are the brain’s messengers, enabling electrical or chemical signals to pass from axon and dendrite.
The BBP effectively served as Markram’s proposal dry run for the HBP, and has now been subsumed by it. One of the BBP’s scientists, Marc-Oliver Gewaltig, explains there are anything between 1,000 and 5,000 synapses per neuron. And in a rat’s cortical column there are perhaps 30,000 neurons. That increases to perhaps 100,000 neurons in each human cortical column. Each cortical column measures about 2mm high and 0.3mm wide. The human cortex – our brain’s wrinkly, grey-green surface – consists of as many as two million cortical columns.
The complexity of that system is unlike any other network that we know of. Mapping it through sampling and physical experiment has proved beyond us. 'It’s impossible,’ Markram says. 'We cannot experimentally map out the brain. It’s just too big. In a piece of the brain the size of a pinhead there are 3,000 pathways like a city with 3,000 streets. It’s taken us 20 years to map 22 of these pathways.’ His solution? Model it. Use the data we have to predict what lies inside the holes in our knowledge and then do experiments to test those predictions. 'To build this thing you’ve got to connect the dots.’ Now he has got €1 billion to do just that.
The award is hugely significant. It guarantees that a vast network of researchers will be stitched together across Europe to create a neuroscience team with a huge range of expertise in disciplines ranging from medicine to computing. It is what Aebischer calls 'info-bio-nano-cogno-science’.
Vast quantities of data and experimental results will be collated and shared. With each bit of information that comes in, more of Markram’s dots will get joined. If he succeeds in joining all those dots, and his 10 years of FET-funded research ends with something approaching a human brain on a computer, Markram will surely win celebrity and acclaim. But the promised benefits to mankind will be considerably greater.
The most significant of these would come in medicine, with the potential to find cures for brain diseases – from autism to Alzheimer’s – that in the EU currently cost hundreds of billions each year to treat. Markram is careful not to guarantee cures. But he does guarantee a computer model on which researchers and pharmaceutical firms can, for example, test out their new ideas, so that they can rule out drugs that won’t work before they get to expensive clinical trials.
In this way he hopes to revolutionise the whole research process – rapidly accelerating it at a time when drug discovery for brain disease is currently so poor that, researchers say, we can’t expect to see any new medicines in our lifetimes. The ability to see the detail of how such diseases manifest themselves deep within the cortex could change that, proving the key to countering their ravages. 'It’s a huge leap forward,’ Markram promises.
To achieve that leap forward, however, will require a great leap forward in computing. Steve Furber, one of the designers of the BBC Micro (which brought computing into homes and schools in the early 1980s) and the hugely successful ARM microprocessor, is working to solve the problem.
There are two issues – processing power, and electrical power. The brain performs innumerable calculations simultaneously, dwarfing the calculating power of our best supercomputers. And it does that using about 20W of power.
'The best estimates of the power we’d need for a computer to model the brain currently come out at about 20GW,’ Furber, now a professor of computer engineering at the University of Manchester, says. The generating capacity of Britain’s biggest nuclear power station, Heysham in Lancashire, is about 2.4GW. 'Our current high-end supercomputers fill an aircraft hangar, and those machines are 1000th of the capability required to model human brains.’
The answer, he speculates, is to abandon the traditional 'architecture’ that has governed the circuit boards inside our computers since the 1950s, and which keeps memory in one place and processing power in another. He is looking to the brain, which combines memory and processing in its cortical columns, for inspiration. For example, current computers are bad at sending vast numbers of very small pieces of information around – which is characteristic of the flow of information in the brain. In the brain each neuron cell performs perhaps 100 operations per second – a tiny number in supercomputing terms.
But with each neuron performing those operations in parallel with every other neuron, the overall processing power is phenomenal. If that 'architecture’ could be replicated in what people like Furber call a 'neuromorphic’ silicon chip, it would herald a revolution in computing. Energy use would fall away dramatically, processing power would rise. That is why, despite the gulf between the best efforts of our current supercomputers and the brain, Furber is convinced that 'it will not be lack of computing power that will prevent the HBP from being a success’.
Furber’s attempts to build a computer that apes the workings of the human brain have resulted in the SpiNNaker (Spiking Neural Network Architecture) project at Manchester. It was born, in 2005, of his frustration that, despite computers existing for decades, 'we still basically could not work out how to endow our machines with anything we would recognise as intelligent behaviour.’
Today, Furber says, SpiNNaker is lighter, cheaper and 'better able to support real-time models’ than traditional supercomputers. He is confident that 'it will play a significant role in this overall quest to understand how the brain processes information.’
Still, anyone with a smartphone can recognise his basic grumble with the state of technological 'intelligence’ today. Tasks that humans perform with ease, such as language recognition, often seem beyond phones’ 'personal assistants’, such as the iPhone’s Siri.
'We would like to see machines become a bit more intelligent,’ Furber says. It would make them easier to use. 'Every time I buy a new computer I have to learn how to use it,’ he says. 'Whereas if you appoint a new PA it’s their job to learn how to work with you. My view is that computers should meet users closer to the middle, and that requires some kind of model of how the user works.’
For example, already in development at EPFL are prosthetics and wheelchairs that can be controlled directly by users’ thoughts – without the need for joysticks or other controllers, which are beyond the use of, say, those afflicted by locked-in syndrome. At EPFL Prof José Millán has developed a brain-machine interface that, after four afternoon sessions, has allowed patients to control wheelchairs in this manner.
For critics all the way back to Jack Good, there is little comfort in the prospect of machines that are 'a bit more intelligent’. For them, each extra grain of intelligence takes us closer to what Huw Price calls that 'Pandora’s box moment with artificial intelligence that, if missed, could be catastrophic’. Furber recognises that 'one has to bear these sorts of long-term threats in mind’, but considers the prospect 'so far off that… it seems a negligible threat at this stage’.
Nevertheless, HBP has a fully-fledged 'ethics’ division, whose aim will be to raise public awareness of the potential dramatic technological upheavals at hand. It is likely to be modelled on the ethical, legal and social issues wing of the Human Genome Project, which absorbed three to five per cent of that project’s total budget, and will use radio and television programmes, conferences, publications and websites 'to encourage well-informed public debate’.
As Richard Walker, part of the ethics division at HBP, puts it, 'We have a clear idea about our limitations. We are not playing God. Or we are just a very minor god at the moment – a very long way from being omnipotent.’
The development of neuromorphic computing systems 'with the ability to learn new tasks without explicit programming’ is, according to the project’s report to the EU, a 'key goal of the HBP’, with potential applications for 'domestic and industrial robots, vehicles and policing, the monitoring of large-scale telecommunications, power distribution and transport networks’.
The report acknowledges that such potential could also be turned to 'implement new systems of mass surveillance and new weaponry… In summary, the HBP offers many benefits but also has risks.’ That is why, Walker says, three per cent of the HBP’s funding is spent by its ethics division. 'We want to have a big public debate. That might not make the problem go away – but at least the public should be aware.’
There is still a risk that the whole project is a giant, expensive flop.
At a Swiss Academy of Sciences meeting in Bern on January 20, only eight days before the FET prize was announced, some were warning that HBP might fail to live up to its hype. 'It’s crap,’ one neuro-scientist told Nature. Other sceptics pointed to the achievements of the Blue Brain Project, which has, after six years, simulated only a tiny slice of rat cortex – a slice that is not connected to sensory inputs. Kevan Martin, a co-director of the Institute for Neuroinformatics in Zurich, worries that with its scope and funding, HBP will 'starve’ other neuroscience research of resources. In an email published in Nature, Martin acknowledged that passionate men such as Markram are needed to drive science forward, then asked, 'But what if they’re passionately wrong?’
On first impressions, self-doubt is not a quality that seems to hamper Henry Markram. And although he has never said so explicitly, it is possible that he has a personal motivation. According to a New Scientist article of 2008, his own son is 'borderline autistic’. Speaking at events, Markram frequently addresses autism. When he does, he refers to the Intense World Theory, which suggests that autism is a response to brain overload. In 10 years’ time, Markram may be able to see the 'intense world’ in action at the level of individual neurons in his computer model of the brain. Then he could do something about it. For, ultimately, the goal of the HBP is to examine the interaction between the brain and the external world – patterns of consciousness and emotion that propel the aims of the project from neuroscience into philosophy.
'We can see patterns of neural activity which are present when someone is conscious of something and not present when they aren’t,’ Walker says. 'We want to build a machine that has that pattern of electrical activity. Is that consciousness? That is an extremely difficult question. Emotions are also associated with patterns of neural activity. If I can represent the patterns of activity associated with being hot, I can probably do the same with feeling sad. But whether that actually ends up with hotness or sadness is up for debate.’
In a lecture in Oxford in 2009 Markram shared with his audience a theory of how the brain works. He suggested that 'the brain builds a version of the universe and projects this version of the universe like a bubble all around us. So I can say with some certainty, “I think therefore I am.” But I cannot say, “You think therefore you are,” because you are within my perceptual bubble. We can speculate and philosophise, but in the next hundred years we won’t have to. We can ask very concrete questions: can the brain build such a perception? Is it capable of doing it?’
The key, he says, is to see each cortical column like a key on a grand piano – the million-key grand piano that is our neo-cortex. 'You stimulate the neo-cortex, it produces a symphony. But it’s not just a symphony of perception. It’s a symphony of your universe. Your reality.’
His dream is nothing less than to play that symphony inside a machine. In his mind’s eye, he can see the machine he wants to build. Now, he says, 'Let’s switch it on. And to switch it on you have to make it come alive.’