For years, Henry Markram has claimed that he can simulate the human brain in a computer within a decade. On 23 January 2013, the European Commission told him to prove it. His ambitious Human Brain Project (HBP) won one of two ceiling-shattering grants from the EC to the tune of a billion euros, ending a two-year contest against several other grandiose projects. Can he now deliver? Is it even possible to build a computer simulation of the most powerful computer in the world – the 1.4-kg (3 lb) cluster of 86 billion neurons that sits inside our skulls?
The very idea has many neuroscientists in an uproar, and the HBP’s substantial budget, awarded at a tumultuous time for research funding, is not helping. The common refrain is that the brain is just too complicated to simulate, and our understanding of it is at too primordial a stage.
Then, there’s Markram’s strategy. Neuroscientists have built computer simulations of neurons since the 1950s, but the vast majority treat these cells as single abstract points. Markram says he wants to build the cells as they are – gloriously detailed branching networks, full of active genes and electrical activity. He wants to simulate them down to their ion channels – the molecular gates that allow neurons to build up a voltage by shuttling charged particles in and out of their membrane borders. He wants to represent the genes that switch on and off inside them. He wants to simulate the 3,000 or so synapses that allow neurons to communicate with their neighbours.
Erin McKiernan, who builds computer models of single neurons, is a fan of this bottom-up approach. “Really understanding what’s happening at a fundamental level and building up – I generally agree with that,” she says. “But I tend to disagree with the time frame. [Markram] said that in 10 years, we could have a fully simulated brain, but I don’t think that’ll happen.”
Even building McKiernan’s single-neuron models is a fiendishly complicated task. “For many neurons, we don’t understand well the complement of ion channels within them, how they work together to produce electrical activity, how they change over development or injury,” she says. “At the next level, we have even less knowledge about how these cells connect, or how they’re constantly reaching out, retracting or changing their strength.” It’s ignorance all the way down.
“For sure, what we have is a tiny, tiny fraction of what we need,” says Markram. Worse still, experimentally mapping out every molecule, cell and connection is completely unfeasible in terms of cost, technical requirements and motivation. But he argues that building a unified model is the only way to unite our knowledge, and to start filling in the gaps in a focused way. By putting it all together, we can use what we know to predict what we don’t, and to refine everything on the fly as new insights come in.
The crucial piece of information, and the one Markram’s team is devoting the most time towards, is a complete inventory of which genes are active in which neurons. Neurons aren’t all the same – they come in a variety of types that perform different roles and deploy different genes. Once Markram has the full list – the so-called “single-cell transcriptome” – he is confident that he can use it to deduce the blend of different neurons in various parts of the brain, recreate the electrical behaviour of each type of cell, or even simulate how a neuron’s branches would grow from scratch. “We’re discovering biological principles that are putting the brain together,” he says.
For over two decades, his team have teased out the basic details of a rat’s neurons, and produced a virtual set of cylindrical brain slices called cortical columns. The current simulation has 100 of these columns, and each has around 10,000 neurons – less than 2% of a rat’s brain and just over 0.001% of ours. “You have to practice this first with rodents so you’re confident that the rules apply, and do spot checks to show that these rules can transfer to humans,” he says.
Eugene Izhikevich from the Brain Corporation, who helped to build a model with 100 billion neurons, is convinced that we should be able to build a network with all the anatomy and connectivity of a real brain. An expert could slice through it and not tell the difference. “It’d be like a Turing test for how close the model would be to the human brain,” he says.
But that would be a fantastic simulation of a dead brain in an empty vat. A living one pulses with electrical activity – small-scale currents that travel along neurons, and large waves that pass across entire lobes. Real brains live inside bodies and interact with environments. If we could simulate this dynamism, what would emerge? Learning? Intelligence? Consciousness?
“People think I want to build this magical model that will eventually speak or do something interesting,” says Markram. “I know I’m partially to blame for it – in a TED lecture, you have to speak in a very general way. But what it will do is secondary. We’re not trying to make a machine behave like a human. We’re trying to organise the data.”
That worries neuroscientist Chris Eliasmith from the University of Waterloo in Ontario, Canada. “The project is impressive but might leave people baffled that someone would spend a lot of time and effort building something that doesn’t do anything,” he says. Markram’s isn’t the only project to do this. Last November, IBM presented a brain simulation called SyNAPSE, which includes 530 billion neurons with 100 trillion synapses connecting them, and does... not very much. It’s basically a big computer. It still needs to be programmed. “Markram would complain that those neurons aren’t realistic enough, but throwing a ton of neurons together and approximately wiring them according to biology isn’t going to bridge this gap,” says Eliasmith.
Eliasmith has taken a completely different approach. He is putting function first. Last November, he unveiled a model called Spaun, which simulates a relatively paltry 2.5 million neurons but shows behaviour. It still simulates the physiology and wiring of the individual neurons, but organises them according to what we know about the brain’s architecture. It’s a top-down model, as well as a bottom-up one, and sets the benchmark for brain simulations that actually do something. It can recognise and copy lists of numbers, carry out simple arithmetic, and solve basic reasoning problems. It even makes errors in the same way we do – for example, it’s more likely to remember items at the start and end of a list.
But the point of Spaun is not to build an artificial brain either. It’s a test-bed for neuroscience – a platform that we can use to understand how the brain works. Does Region X control Function Y? Build it and see if that’s true. If you knock out Region X, will Spaun’s mental abilities suffer in a predictable way? Try it.
This kind of experiment will be hard to do with the HBP’s bottom-up architecture. Even if that simulation shows properties like intelligence, it will be difficult to understand where those came from. It won’t be a simple matter of tweaking one part of the simulation and seeing what happens. If you are trying to understand the brain and you do a really good simulation, the problem is that you end up with... the brain. And the brain is very complicated.
Besides, Izhikevich points out that technology is quickly outpacing many of the abilities that our brains are good at. “I can do arithmetic better on a calculator. A computer can play chess better than you,” he says. By the time a brain simulation is sophisticated enough to reproduce brain’s full repertoire of behaviour, other technologies will be able to do the same things faster and better, and “the problem won’t be interesting anymore,” says Izhikevich.
So, simulating a brain isn’t a goal in itself. It’s an end to some means. It’s a way of organising tools, experts, and data. “Walking the path is the most important part,” says Izhikevich.