Emulating the brain on a chip

Discover Magazine has an article on an innovative project to create silicon chips which work like neurons. If you’re thinking these are standard digital chips that run neural network software you’d be wrong, they’re part-analogue devices that are specifically built to emulate the physical operation of brain cells.

The article riffs on the work of neuroscientist Kwabena Boahen who leads the ‘Brains in Silicon’ project.

If you’re not familiar with the difference between analogue and digital calculation it’s worth just briefly getting to grips with it so you can see how revolutionary this project is.

Most computer chips are digital. They encode numbers as lists of 0s and 1s because they are made up of millions of transistors which can switch on (a ‘1’) and off (a ‘0’). The chip can then do operations or maths on the numbers, by flipping the switches, depending on what functions are built-in and how software makes use of them.

So if you wanted to calculate, lets say, how fast a crowd of people walk through a door, you would need to enter numbers for the size of the door, how fast the people are walking, the amount of interference caused by jostling and crowding and your mathematical formulae which ties it all together. The chip would do the calculation, and you would get your answer.

An analogue calculation is more more like a simulation. For example, you might find that ball bearings and a funnel give you a good approximation of the answer. You just change the size of the funnel, the number of ball bearings and the pressure from behind and you just observe what is happening to get the answer. It might not be as pinpoint accurate, but its much easier to build and run.

The traditional approach to artificial neural networks is the first. Each virtual neuron is a mathematical simulation of the electrical and chemical processes and how it influences other virtual neurons. This needs huge amounts of calculations because each of the simulated neurons is mathematically complex and any change means every connected neuron needs also to be recalculated.

This is the approach taken by the Blue Brain Project and it is no accident that they use one of the world’s biggest supercomputers to run the simulation.

This is where Boahen’s project comes in. While the traditional digital approach is very accurate, its very time and energy intensive. While the Blue Brain project needs a warehouse of tech to support it, the actual noisy error-prone brain runs in the space of a bag of Doritos.

So instead of going for the pinpoint accuracy of digital simulation, Boahen has created chips that are an analogue simulation, or really, an analogue emulation of neurons.

As neurons use electrical impulses, much of their function can be described as electrical circuits. In fact, the Hodgkin-Huxley model of the neuron can be drawn as an electrical circuit.

So instead of writing mathematical equations to simulate the circuit and then getting a chip to do the digital calculations, you could just build the circuit. Using the circuit would tell you exactly how the neuron would behave.

Complete neurons are more complex than the simple Hodgkin–Huxley circuit (which just aims to describe the electrical action potential signal) but the same approach applies. Instead of building a chip to run digital simulations of circuits, just build the circuits. The result is noisy, dirty but fast, very low power and good enough Рjust like the human brain.

We covered Boahen’s work back in 2007 and there’s a great talk he did which introduces the project, but the Discover article is a great update on the research which has the potential to turn neurally inspired computing on its head.

It also has loads of background information and is a great introduction to how the brain deals with its noisy and surprisingly unreliable neurons.

Link to Discover article on brain chips.