Tuesday, 22 March 2011

“Brains on a chip” not the latest recipe from Heston Blumenthal.

Scientists often indulge in international travel for no better reason than to all be in the same place at the same time so they can work together, and disagree about everything. Conferences are a great way to be reminded not only of the perpetual disagreement among experts, but that those of us trying to understand the brain are short of answers. In fact, frustratingly, even the questions are often beyond us.

Of course progress has been made. (It was once believed that the brain was a sort of air-cooled radiator designed to carry heat away from the heart.) In particular, our ability to map centres of activity in the brain gets more sophisticated daily. But for many of us this sort of ‘geographical’ description, however impressive, is a distraction from the inconvenient fact that we have only a sketchy appreciation of how the brain does stuff, even if we are much better at discovering where the stuff is being done.

Imagine you are a Martian astronomer studying the Earth with nothing other than a heat-sensitive telescope with a resolution of around a kilometre. There is much you could learn: what earth-dwellers call day and night would be clear temporal patterns; the tides, towns and roads, would all show up in your data. As your skills developed you could pick up differences between weekdays and weekends, and although you wouldn’t know why this happened you would congratulate yourself on the predictive power of the ‘5+2 rule’ (which you might even be lucky enough to have named after you). Eventually you would even spot subtleties in the direction of the energy flow: into the cities in the morning and out of them in the evening. In these terms, the problem facing neuroscientists is equivalent to our Martian friend using data from his telescope to work out the rules of poker. Not a hope.

So, how to get to grips with the really hard stuff. One approach currently being discussed is to build something that works like a brain, and, more significantly, to build it from silicon brain-like components. This is called “neuromorphic engineering” a useful description even if (in the light of what I said earlier about the perpetual disagreement) not everyone agrees on exactly what it means.

The brains and nervous systems of living things consist of a number of interconnected specialised cells called neurons. Surprisingly, you don’t need very many of these to build a working, living, behaving creature: the humble Caenorhabditis elegans, an unprepossessing worm that lives in soil, manages on exactly 302 neurons. Despite this apparent simplicity its behaviour is, in most ways, much more sophisticated than the best we can manage in machines whose ‘intelligence’ is embodied in high-powered computer systems. How is this done in so few neurons?

Surely if we could determine exactly how these 302 cells are connected and build a mathematical description of how each cell works, then we could build, or at least simulate, a robot worm? Neither I nor anyone else in the field is holding their breath. This is despite the fact that the pattern of connections for the worm’s entire nervous system are known. Human brains have considerably more neurons than the humble C. elegans, around a hundred billion is the usual answer, so we can safely conclude that neuroscientists are in it for the long haul.

Neuromorphic engineers have opened up a new front in this campaign. Giacomo Indiveri is one of the leading exponents of the art of designing silicon chips that don’t execute programs but act like biological systems. I have been working with his group and he told me that the devices we are working on can already ‘exhibit behaviours that model simple properties of neural circuits in the brain’ exhibiting ‘very simple life-like behaviours’. However, ‘they cannot definitely exhibit the full range . . . that even simple insects exhibit in their daily routine tasks’. Giacomo, along with many others in the field, is excited by this, and excited by the possibility that brain-like devices could one day outperform the technology in conventional robotics.

But for me the real promise of this work is more profound. In contrast with the conventional robotics approach, we are starting with the silicon neuron, a component which, like the real neuron, is very unpromising from an engineering point of view. Any progress we make towards building a working device is hard won, and could provide us with a deep and genuine insight into how the brain actually does the stuff.

This is not the end of the story. Neurons only make up about 10% of the volume of your brain. The rest consists mostly of cells called `glia’ (from the Latin meaning `glue’) that were once thought to be there simply to hold everything together. No such luck.