Building a Brain

If you really want to understand how the brain works, you have to build one, according to Kevan Martin, who is doing just that at the Institute for Neuroinformatics in Zurich. Martin and his colleagues analyse the workings of individual brain cells, trace the circuits they form, and build silicon microchips that work exactly like brain cells linked together to make networks that process information in the same way as the brain does.

The more conventional approach to understanding how networks of neurons work is to write a computer program that calculates how the cells work, and simulates the effect of connecting them together. Simulations are important because it is impossible to investigate both the properties of the cells and those of the circuits they form. Each cell calculates its output from hundreds of different input signals, and sends its output signal to hundreds of other cells.

The signals in brain cells can only be analysed in one or two cells at a time.Tracing networks of cells to show what kinds of circuits they form can only be done on specially prepared (dead) tissue. So simulation provides a way of working out how networks of cells behave, and how an individual cell’s behaviour may contribute to the function of the whole circuit.

Martin insists that computer simulations don’t tell you enough. “Imagine that the Wright brothers had been able to use a computer to simulate powered flight in 1903.” He says. “It wouldn’t have helped. It’s only when you build a plane that you discover how difficult it is to fly one. Early planes used to stall and crash. The only reason the Wrights didn’t fail is that they put the tail on the front, so the plane didn’t stall.”

Martin aims to build a seeing machine rather than a flying machine. He wants to understand how a part of the brain called the visual cortex processes the information that comes from the eyes. A silicon version of the retina, the sheet of nerve cells that processes information in the eye before it is sent to the brain, was designed by Carver Mead of the California Institute of Technology and built in 1990 by Misha Mahowald [correct], who now works with Martin in Zurich. “The retinal circuitry was already fairly well worked out. The trick was to fabricate it using digital chip technology (known as VLSI) and to figure out how to make the circuits work.” says Martin.

The voltages in digital VLSI chips flip instantaneously between 0 and 5 volts. Digital computers represent different numbers by the same voltage in different parts of the circuit. In analogue computers, like nerve cells, the voltage is proportional to the number it represents. Voltages must change smoothly and continuously over the neurone’s operating range of about ten millivolts. Mead discovered that it was possible to make digital VLSI chips work in analogue mode by restricting their voltage range to about half a millivolt. This makes it possible to design “silicon neurones” that mimic exactly the operation of any conceivable brain cell.

One of the first problems with silicon neurones is that it’s impossible to standardise the components. They all work slightly differently. “But the same thing happens in a real brain” says Martin. All neurones are slightly different, and the brain’s circuits are designed so that they can compensate for the differences. This is why Martin’s work has a strong biological component too. “We are studying the brain in fine detail to figure out what connections are needed in the silicon” he says.

Wiring up the circuits of silicon neurones presents a special problem. “The brain can run a wire from any cell to any other cell. You can’t do that on a chip because it’s flat. You just get too many wires.” Says Martin. The solution is to use an addressed bus to connect all the neurones together. Each neurone labels its output and sends it to the bus; a digital signal processor sends it to all the neurones in the ciircuit that should receive it as input. Martin’s group are working to build a circuit containing 10,000 cortical neurones, about as many as are contained in 0.5 square millimetres of visual cortex.

A surprising benefit of silicon neurones is that they use very little power. “If you had to build your brain in conventional digital technology it would consume about as much power as a small town” says Martin. Silicon neurones are about 10,000 times more efficient, although they still use about 100,000 times as much power as real neurones. NASA are discussing the possibility of using analogue VLSI technology for remote sensing devices on Mars.

But Martin’s goal is to understand how the brain works. “The driving idea is that all of neocortex (about 80% of the brain) is built from repeated units of the same basic circuit…. If we can define its connectivity and build it in AVLSI then maybe we’ll know why people are so clever” he says.