Computer model of a single neocortical column from a rat's brain (Photo: IBM)
Beside the fact that such news are impressive, I liked what the scientist had to say to the reporter and what the reporter extracted from this complex information. One can appreciate the accuracy and the honesty of this article when it is compared to another article on Blue Brain I read in Spiegel a while ago (leave the Spiegel article for the end and you will see what I mean, you will be able to read it with a more critical eye thanks to the present Guardian article). There is nothing sensationalist about Clint Witchalls' article, and in addition he asks the right questions. Still, the title is catchy and I bet it is the editor's touch on the article.
Blue Brain is not about Artificial Intelligence taken in its direct sense, it is about adopting a smart approach to a very complex issue, namely modelisation. Experimental scientists who are hard empiricists abhorre modelisation. You can tell from Steven Rose's comment at the end of this article and I think he is quite right, for the moment. But I think we cannot understand everything from scratch, from trial and error. We have to figure out some way to anticipate what we are trying to understand and modelisation does just that. And nothing can stop us from testing our anticipations. Models, artificial models in experimental science, are just practical material hypotheses. We don't have to be afraid of them or suspicious. And as other hypotheses, they may end up wrong. It is not because we were able to construct a model, a material thing, out of a hypothesis, that the hypothesis becomes instantly true, embodied in a certain way in our own material creation. Sailors before Copernicus used to navigate with totally false charts of the sky, and still, some of them I suppose managed to find their way in the sea. That does not mean that theoretical hypotheses about the stars and the sky for these charts were true. Similarly, the modelisation of a Rat Neo-cortical column, even though it may seem small in scale in modeling the brain, and unfaithful to the original biological model, is a major breakthrough in Neuroscience, both at the practical, conceptual and epistemological levels. The model may turn out to be wrong, and nothing like the real brain. However, along the way we would have learned and amassed a tremendous amount of data otherwise unaccessible to scientists.
In a laboratory in Switzerland, a group of neuroscientists is developing a mammalian brain - in silicon. The researchers at the Ecole Polytechnique Fédérale de Lausanne (EPFL), in collaboration with IBM, have just completed the first phase of an ambitious project to reproduce a fully functioning brain on a supercomputer. By strange coincidence, their lab happens to lie on the same shores of Lake Geneva where Mary Shelley dreamt up her creation, Dr Frankenstein.
In June 2005, Henry Markram, director of the Blue Brain project, announced his intention to build a human brain using one of the most powerful supercomputers in the world. "The critics were unbelievable," recalls Markram. "Everybody thought we were crazy. Even the most eminent computational neuroscientists and theoreticians said the project would fail."
Some of Markram's peers said there simply wasn't enough data available to simulate a human brain. "There is no neuroscientist on the planet that has the authority to say we don't understand enough," says Markram. "We all know a tiny slice. Nobody even knows how much we know."
Markram was not dissuaded by the negative reaction to his announcement. Two years on, he has already developed a computer simulation of the neocortical column - the basic building block of the neocortex, the higher functioning part of our brains - of a two-week-old rat, and it behaves exactly like its biological counterpart. It's something quite beautiful when you watch it pulse on the giant 3D screens the researchers have constructed.
The neocortical column is the most recently evolved part of our brain and is responsible for such things as reasoning and self-awareness. It was a quantum leap in evolution. The human brain contains a thousand times more neocortical columns than a rat's brain, but there is very little difference, biologically speaking, between a rat's brain and our own. Build one column, and you can effectively build the entire neocortex - if you have the computational power.
Although a neocortical column is only 2 millimetres long and half a millimetre in diameter, it contains 10,000 neurons and 30m synapses. The machine that simulates this column is an IBM Blue Gene/L supercomputer is capable of speeds of 18.7 trillion calculations per second. It has 8,000 processors and is one of the most powerful supercomputers in the world.
Markram believes that with the state of technology today, it is possible to build an entire rat's neocortex, which is the next phase of the Blue Brain project, due to begin next year. From there, it's cats, then monkeys and finally, a human brain.
Markram is banking on Moore's law holding steady, as a computer with the power of the human brain, using today's technology, would take up several football pitches and run up an electricity bill of $3bn a year. But by the time Markram gets around to mimicking a full human brain, computing will have moved on.
Modelling the future
Modelling seems to be the way forward for neuroscience. Each year, there are about 35,000 neuroscience papers published - and the number of papers being published is increasing at a rate of between 20% and 30% a year. Most neuroscientists only get to read about 100 of these papers a year, if they're lucky. Pouring all of this knowledge into Blue Brain seems an obvious way to use and preserve it.
Markram, a 44-year-old South African, first became interested in recording the electrophysiology of neurons when he was at the Max Planck Institute in Germany. He was recording two neurons and he saw them communicate. "I thought, my God, this is incredible, you can actually capture neurons communicating," he says. "Then I wanted to find out how they all communicated, so I started to map the whole circuit. It took 15 years." Markram describes the data he has collected over the past decade and a half as "too boring to be published".
The model is there to unify the data and test that it works. A neurobiologist who wants to test a certain theory of how a specific brain function, such as memory retention and retrieval, works can use Blue Brain to do so. The model will be open to the entire world's research community.
Simulation-based research becomes possible when you have a critical power of computation. Today, every commercial aircraft that is built started life as a simulation. Even cameras are simulated before they're built. In physics, we don't let off nuclear weapons any more, we just use simulations.
"We don't use simulation in life sciences because biology requires the most powerful computers," says Markram. "We do experiments on animals, but that is going to change in the near future and this project will drive that change."
One thing Markram is keen to stress is that this isn't another artificial intelligence (AI) system. "We're not looking for the brain of a robot," he says. "You can get an engineer to do that. They are much better at it and they can do it really quickly. But in the end, it [Blue Brain] will probably be much better. If we build it right, it should speak."
However, Markram is not holding his breath, waiting for some emergent consciousness to arise from the silicon brain. What he is after is something more prosaic, but also a lot more useful than a talking machine. By understanding the function of the brain, we can also begin to understand its dysfunction.
Disorders such as depression, schizophrenia and dementia are the price we pay for having complicated brains. "We don't understand what goes wrong inside those circuits," says Markram. "We're still in empirical medicine. If a drug compound works: good. If not, we try another one." Blue Brain could accelerate experimentation tremendously. It will be much more efficient than wet-lab experiments and it will reduce animal experimentation.
However, Steven Rose, emeritus professor of biology at the Open University, is sceptical that a biologically accurate model of the entire human brain can be built, given our current state of knowledge and technology. The integration between the different regions of the brain is just too complex to recreate on a computer simulation. "I'm not against people playing with models," says Rose, "but the idea that you can use it for anything very sophisticated as opposed to looking at real animals with real behaviour at the moment seems to me to be pie in the sky."
Rose warns against underestimating the difficulties that still remain. Then, rather grudgingly, he admits that the Blue Brain project is impressive. "Impressive but modest," he adds. Clearly, Markram still has some doubters to win over.