Those of a certain age will remember Acorn’s BBC Micro computer, designed for the BBC’s Computer Literacy Project. Professor Steve Furber, who developed the computer’s microprocessor, ARM, which today is used most notably in smart phones is once again working at the cutting edge of computing – leading a project called SpiNNaker (Spiking Neural Network Architecture) which aims to build a new type of computer that can model the human brain. Below he talks about SpiNNAker and what it could mean to artificial intelligence, science and computing…
One question that scientists get asked a lot is; how far away we are from creating artificial intelligence? The answer depends of course on what we mean by intelligence but if we are talking about something that is an artificial version of man then the answer is unfortunately (or perhaps fortunately if you grew up watching Terminator) we aren’t even close.
The main reason for that is the human brain. While we are now able to grow ears on mice and create artificial internal organs we are a long way from building a brain. The average brain contains around 100 million neurons connected by a quadrillion constantly changing synapses. Although it uses less energy than is needed to power the average light bulb, if you attempted to recreate this with the computing power available today you would need a nuclear power plant to run it.
The ability to build a model of the human brain would be a milestone in terms of improving our understanding of how we operate. It would be an amazing tool for scientists looking at brain diseases and neurological conditions like depression as well as giving us an insight into how we learn and the connections between brain and body. From a technical viewpoint – understanding how the brain does so much while consuming so little energy would pave the way for ultra-fast, energy efficient chips.
There are a number of projects scattered across the globe that are dedicated to modelling the human brain. To date however these have been based on analogue rather than standard digital circuits. While very fast and very efficient they are also rather inflexible. The models have to be built and rebuilt every time you need to make a change. Our challenge at SpiNNaker (Spiking Neural Network Architecture) over the next 18 months is to create the first low power, large-scale digital model of the brain. The big difference with digital circuits is that they will allow us to meet the speeds of a biological brain but have the programming flexibility of a supercomputer.
The major challenges for us are two-fold: firstly we need the processing power. This model will require one million ARM chips – the same chips that you find in a standard smartphone today (and even then it will only have the complexity of one percent of the human brain). Secondly – and perhaps more significantly, is the fact that the human brain doesn’t process data in the same way that a standard supercomputer does. Traditionally computers are linear. They complete one process after another. The brain however handles multiple tasks simultaneously. In the technology world we call this parallelism.
Many supercomputers operate like this but to do that they need to be huge because they require considerable amounts of power. Add to this the complication that connections within computers are fixed but within the brain they are constantly changing and you see why modelling the human brain is such an incredibly complex task.
The key for us at the SpiNNaker team is in the fusion of technology and medical science. The only way we can begin to meet this challenge is by working with researchers who study the way the brain behaves. Already huge steps have been taken by researchers like those at the Blue Brain Project in Lausanne who have combined the power of a supercomputer with data collected from very detailed studies of brain tissue to simulate the way neurons behave in small sections of a rat’s brain.
The challenge is a huge one but we have already created a prototype that will eventually be scaled up into something that neuroscientists and psychologists can use to test their hypotheses on neural patterns. We hope that this will bring us one step closer to unravelling the mysteries of brain and help us find ways to counter some of what happens when it goes wrong.