This million-core supercomputer inspired by the human brain breaks all the rules

For all their fleshly failings, human brains are the model that computer engineers have always sought to emulate: huge processing power that's both surprisingly energy efficient, and available in a tiny form factor. But late last year, in an unprepossessing former metal works in Manchester, one machine became the closest thing to an artificial human brain there is.

The one-million core SpiNNaker -- short for Spiking Neural Network Architecture -- is the culmination of decades of work and millions of pounds of investment. The result: a massively parallel supercomputer designed to mimic the workings of the human brain, which it's hoped will give neuroscientists a new understanding of how the mind works and open up new avenues of medical research.

The genesis of the project lies in the late 1990s, with the work of Steve Furber, now professor of computer engineering at the University of Manchester.

"I was beginning to wonder why, with processors having got so much faster, there were things that they found hard to do which, as humans, we find relatively easy," says Furber. He began exploring associative memories and, in trying to solve their difficulties with inexact inputs, turned his attentions towards neural networks.

In 2005, the SpiNNaker project got grant funding, and Furber's engineering group took biological brains as their computing models. SpiNNaker now resides in the University of Manchester's Kilburn building, formerly used to house mainframes in the late 1960s and 1970s.

Historically, the difficulty in making computers that could mimic the brain largely comes down to connectivity. Neurons -- the nerve fibres that travel throughout the body and largely terminate in the brain -- each have thousands of inputs and thousands of outputs. Computing systems struggle with anything on a similar scale. "It was clear that the big problem in building computational models of biological neural networks is getting anywhere close to the degree of connectivity you find in biology," Furber says.

In order to build a system that more closely resembles the human brain, the research group created a novel spiking neural network system-on-a-chip. Spiking neural network architectures take their cue from the way neurons work in the brain: in order to pass a signal from one neuron to another, the voltage of its membrane has to change, and what's known as an action potential has to be generated. The action potential is translated as the spikes in a spiking neural network.

By using this architecture, the team say that SpiNNaker breaks the rules followed by traditional supercomputers because the nodes communicate using these simple messages -- spikes -- that are inherently unreliable. "This break with determinism offers new challenges, but also the potential to discover powerful new principles of massively parallel computation," the team says.

After reaching half a million cores in 2016 as part of EU Human Brain Project, SpiNNAker recently reached one million, enabling it to perform two trillion actions per second and model the action of 200 million neurons in real time.

"SpiNNaker's extremely flexible -- all the models we use of neurons and synapses are little bits of software. If you turned those into hardware, they would be smaller and more efficient, but the reason we use software is there's no real agreement as to what the right model is and different brain regions probably need different models, and software gives us flexibility," says Furber.

SEE: How to implement AI and machine learning (ZDNet special report) | Download the report as a PDF (TechRepublic)

Spiking neural networks are inspired by the brain, it's also hoped that they may in turn shed greater light on the organ they're modelled on. That's because the neural net spikes the way real neurons do and transmits information in the same way.

For example, the SpiNNAker system has already been used to model the basal ganglia, a collection of structures in the brain that helps with movement pattern selection -- when you decide to walk to a chair and then sit down, it's the basal ganglia that send the walking pattern to your legs, start it working, stop it when you reach the chair, and send the sitting down pattern to your lower body instead. It's hoped that by modelling the basal ganglia, neuroscientists will gain greater insight into what they do and how they do it.

And that's not all: a grander ambition is that SpiNNaker can be used to understand the cortex -- the outer layer of the brain that has a critical role in higher functions like speech and decision making. It is being used to model the cortical microcolumn, groups of neurons that run through the cortex.

"When we look at the spike rates, they're about right. The next challenge is to try and test some theories about the function of the microcolumn, because that function is not known... we don't know how any of that works," Furber says.

Further out, it could even be used to predict the effects of medication on the human nervous system, and aid the development of new drugs. However, there are a few major hurdles the scientific community as a whole have to clear before that can happen: there needs to be a better understanding of how drugs affect the complex mix of proteins that make up the 10 14 synapses of the brain. Once that's known, it might be possible to compute how drugs will affect the synapse and then eventually build a system model on SpiNNaker to see how those effects might change behaviour.

SpiNNaker could also be of service in the robotics industry -- whether the robots are real or virtual. The Human Brain Project, which SpiNNaker is now part of, is developing a virtual robotics environment that the University of Manchester is installing on its servers.

That means users from around the globe could set up a robot model and close-couple it to their spiking network brain model in SpiNNaker, says Furber, allowing them to observe its behaviour and control it remotely over the internet.

SpiNNaker is a good target for researchers in robotics, who need mobile, low-power computation. A small SpiNNaker board makes it possible to simulate a network of tens of thousands of spiking neurons, process sensory input and generate motor output, all in real time and in a low-power system.

SEE: Sensor'd enterprise: IoT, ML, and big data (ZDNet special report) | Download the report as a PDF (TechRepublic)

Running at high capacity, its energy consumption can peak at 100KW, but normally, it's more in the region of 1-2KW.

"The power is pretty much proportional to the workload, and actually with neural networks models, you're very hard pressed to get anywhere near the maximum power on whatever subset of the machine you're using. We can run multiple jobs at once; we do that not by timesharing individual boards, but by physically partitioning machines, so that if one job requires 6 boards here and one requires 12 boards there, we can run those together and we basically turn off the communications around the edges of the allocated region so there's no risk of interference," says Furber.

And for those that worry that SpiNNaker's brain-like computing could evolve into a sentient computer, there's probably not too much to fear, at least not yet.

SpiNNaker occupies 11 19-inch rack cabinets, making a machine that's around 5m long, 2m tall, and 1m deep. Currently, it models one percent of the human brain, so a SpiNNaker system that could give a robot human-level cognition would require something of an engineering miracle.

The SpiNNaker team is already looking to the future and, in collaboration with TU Dresden, is already working on a second generation of chip that should provide 10 times the functional density, but it still won't be enough to deliver a robot with humanoid thinking powers.

"There's no sense in which this is a technology that will lead the science fiction walking, talking, intelligent robot", says Furber, "because they'd need a head the size of an aircraft hangar and a nuclear power station attached to it."

PREVIOUS AND RELATED COVERAGE

ANZ bank unpicking neural networks in effort to avoid dangers of deep learning

Rather than pushing a successful proof-of-concept into production, ANZ bank has a number of health checks to work through first.

Nvidia aims to run neural nets faster, more efficiently

As data gets bigger and models grow larger, deep learning is once again "completely gated by hardware." At the VLSI Symposia, Nvidia suggested some ways to address this problem.

Wolfram Research goes for Software 2.0, releases neural net repository

Wolfram, having been into AI before it was cool, now gets a piece of the deep learning hype, in its sui generis way. Where does it stand compared to the competition, and how easy is it to use and integrate Wolfram with the rest of the world?

Symantec develops neural network to thwart cyberattack-induced blackouts

The company is rolling out a device that scans for malware on USB devices to block attacks on IoT and operational technology environments.

Scientists just found a previously unknown part of the human brain (CNET)

A brain cartographer suggests the cluster of cells appears unique to humans and may be responsible for fine motor control.

Symantec develops neural network to thwart cyberattack-induced blackouts (TechRepublic)

The company is rolling out a device that scans for malware on USB devices to block attacks on IoT and operational technology environments.