IBM’s SyNAPSE chip: creating a neural computer like your brain
Computers are pretty good at some things, like number crunching and logic, but terrible at others that are easy for our brains, like understanding images or conversation. And even when researchers train software to do things like recognise faces or cats, as Google has done, it requires hundreds or thousands of computers, consuming vast amounts of energy. The human brain can do so much more and uses only what an energy-saving light bulb would.
IBM hopes to close the gap between humans and machines by creating computer chips modelled on biological brains. The SyNAPSE chip it unveiled in August has a design that mimics the neural networks found inside your head. The chip isn’t great at the stuff traditional computers excel at. But it is far better suited to pattern recognition and processing images, sound and other sensory data. And the chip, which is the size of a postage stamp, requires a strikingly small amount of power – about that of a hearing-aid battery. IBM says its descendants could bring brain-like abilities to everything from smartphones to robots. Early in 2015 the company plans to release a kit to let academics, students and corporate researchers start experimenting with this new kind of computing. It will take the form of a smartphone-sized circuit board with a SyNAPSE chip, and connectors to hook up cameras or other devices. “There is the potential here to really open up a spree of tremendous innovation,” says Dharmendra Modha, chief scientist of brain-inspired computing at IBM. Fresh thinking is needed, he says, because the chip is so different from what has come before.
The SyNAPSE chip, says Modha, could lead to vast advances in the sensory processing ability of smartphones. Imagine a phone, he says, that understands where it is, who is speaking and what you are doing. There are plenty of other potential applications, from self-driving cars that need to fuse information from the world around them to devices to help vision-impaired people navigate through the environment.
Like existing chips, the SyNAPSE chip’s circuits are based on switch-like silicon devices called transistors. But they are arranged in a way that breaks with the basic architecture used in computers for the last 70 years. That architecture forces computers to work on tasks as a series of step-by-step instructions, like a recipe. Brains – and IBM’s new chip – can be more efficient because they can execute many different instructions simultaneously. The SyNAPSE chip has a network of 1m fake neurons connected by 256m fake synapses. That is “roughly on the scale of a bee brain”, says Modha, but way behind the human brain, which has 10bn neurons and 100trn synapses. But his team has recently shown that the chips can be tiled together, offering a route to much greater power.
Other experts have welcomed IBM’s experiment in brain-inspired computing, saying that the septuagenarian model that still powers our latest gadgets probably can’t be pushed much further in terms of computing power or efficiency. However, Murray Shanahan, a professor of cognitive robotics at Imperial College in London, says we are still far from knowing how to programme a chip like IBM’s to do sophisticated tasks, or working out what its limitations are. “We really are just at the beginning of understanding how to do this. There’s a long way to go before we match the abilities of even simple animals,” he says.