Computers

Million-core neuromorphic supercomputer could simulate an entire mouse brain

Million-core neuromorphic supe...
A newly built supercomputer is able to simulate up to a billion neurons in real time, enough to emulate a whole mouse brain and simulate sections of the human brain for pharmaceutical testing
A newly built supercomputer is able to simulate up to a billion neurons in real time, enough to emulate a whole mouse brain and simulate sections of the human brain for pharmaceutical testing
View 3 Images
A newly built supercomputer is able to simulate up to a billion neurons in real time, enough to emulate a whole mouse brain and simulate sections of the human brain for pharmaceutical testing
1/3
A newly built supercomputer is able to simulate up to a billion neurons in real time, enough to emulate a whole mouse brain and simulate sections of the human brain for pharmaceutical testing
The supercomputer features one million processing cores
2/3
The supercomputer features one million processing cores
A detail of the interconnection network between the processing cores
3/3
A detail of the interconnection network between the processing cores

After12 years of work, researchers at the University of Manchester inEngland have completed construction of a "SpiNNaker" (SpikingNeural Network Architecture) supercomputer. It can simulate theinternal workings of up to a billion neurons through a whopping onemillion processing units.

Thehuman brain contains approximately 100 billion neurons, exchangingsignals through hundreds of trillions of synapses. While thesenumbers are imposing, a digital brain simulation needs far more thanraw processing power: rather, what's needed is a radical rethinkingof the standard computer architecture on which most computers arebuilt.

"Neuronsin the brain typically have several thousand inputs; some up toquarter of a million," Prof. Stephen Furber, who conceived and ledthe SpiNNaker project, told us. "So the issue is communication, notcomputation. High-performance computers are good at sending largechunks of data from one place to another very fast, but what neuralmodelling requires is sending very small chunks of data (representinga single spike) from one place to many others, which is quite adifferent communication model."

Theresearchers tackled this problem by devising a massively parallel architecture where each of the million cores is able tosend tiny "packets" of information (up to just 72 bits in size)that are routed to their destinations by an internal communicationnetwork.

A detail of the interconnection network between the processing cores
A detail of the interconnection network between the processing cores

Withthis architecture, the supercomputer should be easily capable ofsimulating the 100 million neurons inside a mouse's brain. Even anad-hoc design, however, isn't nearly enough on its own: to build aproper brain model, you'll also need to get the wiring right.

"Tobuild a mouse brain model we need, in principle, to know every neuronand its connections to every other neuron in the brain," Furbertold New Atlas. "In practice this an infeasible amount of data tocollect, so we have to settle for statistical distributions ofneurons types and statistical connectivity data, so that we canconstruct a statistically representative brain model.

"Suchmodels do now exist, though they are very rough cut in places – theyhave been compared to the first attempts to draw a map of the globe,which had highly variable accuracy and missed out Australiaaltogether as it hadn't been discovered then."

Thoughone-to-one neuronal mapping may not happen anytime soon, even asomewhat rough lay of the land could provide interesting results. Forinstance, researchers could build a computer model of the visualcortex of a mouse, "show" it an image that would be translatedinto a stream of spikes down the optic nerve, and learn much abouthow such a signal is processed by the cortex, even using the outputto control the movement of a virtual mouse or a physical robot.

Furbertells us that the system also has the potential to uncover more abouthow high-level functions such as learning work inside the brain.

"Wealready support a fair amount of work on learning processes at thesynaptic level, including dopamine reinforced plasticity which is abiologically-plausible form of reinforcement learning. But thoughputting such local plasticity rules together into a high-levelbrain-like learning system is possible on SpiNNaker, it is stretchingour understanding to generate such a system that we can then claim'is how the brain learns.'"

Theteam has already used the system to simulate aregion of the brain called the Basal Ganglia, an area affected inParkinson's disease. Indeed, there is potential for this technologyto provide advancements in the medical field, particularly withregard to pharmaceutical testing,though the researchers believe the impact of his research on realpatients could take decades to materialize.

Furberand his colleagues are now working on a second-generation machine,"SpiNNaker2," which uses upgraded silicon technology to deliver10 times the functional density and energy efficiency. This wouldenable, among other things, the creation of a whole insect brainmodel in a system that could fit on top of a drone.

Prof.Forber provides more details on the SpiNNaker project in the videobelow.

SpiNNaker: 1 million core neuromorphic platform

Source:University of Manchester

1 comment
usugo
so, it is an approximation, of an approximation, of an approximation, ...