The way in which systems of neurons in the brain interact as a network could become the inspiration for million-processor parallel computer architectures.
The universal spiking neural network architecture (SpiNNaker) project seeks to mimic the brain’s structure and function to offer massively parallel computing coupled with inbuilt redundancy that would make processors faster and more reliable.
The project, led by Manchester University computing pioneer Prof Steve Furber with contributions from Sheffield, Southampton and Cambridge Universities, has more than £5m funding from the EPSRC’s Large Grant Panel. It is one of a number of projects underway around the world that are seeking to continue the exponential increases in processing power set out in Moore’s Law.
In the brain, communication is achieved through electrical pulses known as spikes between axons —the point where neurons connect.
In a computer, the operation of the billions of neurons in the brain could be simulated using digital processors, and the connectivity by transmitting information packets between large numbers of processors.
Furber, ICL professor of computer engineering in Manchester’s School of Computer Science, began thinking about biologically inspired forms of computer memory 10 years ago.
After some false starts, he did some research into neural-style memories and realised this approach could address problems of robustness and fault tolerance in computer engineering. Combined with advances in processor clock speeds and parallel processing, it could be used to produce powerful computer systems.
Furber hopes the project, which runs until the end of 2013, could lead to a usable million-processor machine. ’In a sense, it’s a slightly artificial goal,’ he said, ’but a million processors could model a billion neurons in real time. This is a very large-scale simulation, but bear in mind a billion neurons is still only one per cent of the human brain. The ultimate goal of producing something that could simulate the human brain is still pretty challenging, even with today’s technology.’
The research will focus on what can be termed middle-layer processing in the brain. The firing of individual neurons and the function of larger areas of the brain are fairly well understood but the processing between the two is not.
’This is because it’s fundamentally difficult to build instruments that can measure the firing patterns of thousands or millions of neurons and insert those into a brain,’ said Furber.
Co-ordinating this middle level is a poorly understood ’neural language’, which governs how populations of neurons communicate information to each other.
’There are many hypotheses around how that happens, from fairly basic to quite sophisticated ways that populations collaborate to communicate information,’ added Furber.
’But we don’t know what the characteristic firing patterns are when you think of your grandmother, for example, how those thoughts are conveyed in terms of neurons firing.’
SpiNNaker aims to build a general-purpose computer platform on which researchers can test hypotheses of how this language might work. ’We intend to support the building of larger-scale systems of neurons than has been possible to date, then look to our colleagues in neuroscience or psychology for hypotheses to use in our model and see how well it supports the evidence,’ said Furber.
Though neurons are complex, he added, much of that complexity has to do with biological functions unrelated to the way they contribute to information processing. The machine will be designed to achieve a level of abstraction sufficient to capture the information processing without losing functionality.
The key to the platform is a bespoke chip design the team is developing alongside processor designer ARM, which has 16 to 20 ARM cores. On-chip technology designed by Manchester spin-out Silistix will interconnect them. Thales will become involved at the application level.
’Neurons inside our heads communicate principally by transmitting spikes, which are electrical impulses,’ said Furber. ’We take in the electrical impulse as an event in the system and we communicate it through a packet switch network. We can’t achieve the level of physical connectivity in the brain, so we use the very high-speed electronics to abstract that and use logical connections. We’ve built bespoke routers that send these packets to the right places.’
When SpiNNaker ends, Furber hopes the team will have a million-processor machine, user-friendly software and some key applications to demonstrate the system’s capabilities. ’It will be quite a big machine — it will require 100,000 chips so it’s not going to be something that fits in your PDA,’ he said.