Making computers more like your brain
Computers beat the human brain at crunching numbers but fall way behind in recognising objects. But that is changing with a new type of neuromorphic, or brain-like, computer chip that could speed up development of robotic eyes and ears. Cathal O’Connell reports.
Your laptop can plot a trajectory to the Moon, yet gets fooled by the “Captcha” – the distorted letters used to tell human from computer in online forms.
It’s the paradox of computers. They leave us in the dust when it comes to calculating but are orders of magnitude worse at recognising images, sounds and objects. What’s more they require vastly more power. A supercomputer simulating the human brain would need the combined power supply of Los Angeles and New York. Our brain gets by on the equivalent of a few grams of chocolate an hour.
Now a new type of “neuromorphic” silicon-chip designed to mimic the brain looks set to give computers more human-like capabilities
The brain chip dubbed “TrueNorth” was unveiled in Science magazine earlier this month. It is the fruit of the SyNAPSE program (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) funded by the US defence research agency DARPA and developed by IBM and an academic partnership between New York’s Cornell University and Israel’s Technion Institute.
Not only can the postage stamp-sized chip recognise the difference between objects almost as well as a human brain, it needs only a tenth of a watt – the power supplied by a hearing aid battery – to do it.
According to IBM’s Dharmendra Modha, who led the project, the new chip has the potential to revolutionise the computer industry.
“I was not there when ENIAC [the first programmable electronic computer] was unveiled but I have a palpable sense that we are at a similar turning point in the history of computing,” he writes on his website.
“It’s a major step,” agrees Stephen Furber, who researches neuromorphic computing at the University of Manchester in the UK. Although the chip itself is big by electronics standards – eight times larger than the typical processor in a smartphone – “they've got a formidable amount of neural stuff on it”.
TrueNorth promises to accelerate the development of everything from robotic eyes and ears to visual aids for the blind. It also helps pave the way to building practical supercomputers with the processing power of a human brain.
Modern computers are still for the most part designed like an old-fashioned office where the secretary is constantly running to open files to work on and then putting them back when finished. That’s the way mathematician John von Neumann designed them in 1945. They follow a sequence of instructions, but the parts of the computer that do the processing and those that store memory are physically separated. As the processor works it must constantly fetch packages of data from the memory, manipulate them, then send them back to be stored. All this back and forth creates a bottleneck and gobbles energy. It also means that the individual chips are working non-stop sending these messages.
By contrast, in a brain, memory and processing occur in the same place – the neuron. Messages travel from one neuron to the next via connections called synapses. To make a memory the neuron simply strengthens the connections of those synapses. Real neurons also have time off. They only fire when the input signals reach a threshold, which saves a lot of energy.
Just one chip can already do things that it currently takes a supercomputer to do, such as identify pedestrians, cyclists and cars on a video feed in real-time.
Each neuron is also connected to 10,000 others making it a masterpiece of parallel processing. Complex tasks can be divided up and done at the same time. For instance, when you look at the room around you different teams of neurons detect the horizontal and vertical lines, other teams process colour, perspective and so on, before the information is passed to other neuronal teams that pull the whole picture together.
Although the signalling between individual neurons is tediously slow compared to a computer chip, the parallel and integrated processing of a brain makes it far superior at sifting through the vast swathes of data needed to orient yourself in the world.
Computer scientists have long tried to emulate brain-like processing using programs known as neural networks. But while they have progressed rapidly in recent times, the networks require vast amounts of computing power.
For instance, in 2012 Google used neural networks to scan through 10 million images extracted from YouTube. With no instructions the network famously identified a dominant image – cats! But that processing feat required the power of 1,000 computers, each sporting 16 processors linked up in a “Google brain”. That large processing requirement owes much to the back-and-forth that goes on in the von Neumann architecture.
TrueNorth, by contrast, doesn’t just rely on brain-like programs, the very architecture of its silicon chip is brain-like.
It has one million electronic “neurons” connected via 256 million “synapses”. As with real neurons the memory and processing functions are embedded in the chip. And the electronic neurons only fire when the signals reach a threshold. As a result, the new chip consumes 1,000 times less power than conventional chips of a similar size.
Just one postage-sized TrueNorth chip can already do what a supercomputer does, such as identify pedestrians, cyclists and cars on a video feed in real-time. But Modha and his team are thinking big.
They next step is to build arrays of neuromorphic chips, layered like the brain’s cerebral cortex and working in parallel to carry four billion neurons and one trillion synapses. Yet these will require a mere 4 kW of power, still equivalent to more than 50 laptops but thousands of times less than conventional architecture would require.
“The ultimate goal of the project is to build, literally, a brain in a box,” Modha says.
Follow Cathal on Twitter @iCatHalo