An artistic view of a neuron looking a bit computerish
Like a neuron, the new chip fires when it reaches a certain threshold

IBM research

There’s nothing quite like the human brain. Today, researchers at IBM unveiled their latest attempt to mimic it: an artificial neuron that switches between crystal and glass-like states as information comes in.

It is designed to better handle huge volumes of data at a fraction of the energy cost of conventional chips. “The challenge here is to receive data that is increasingly big and complex, and extract useful knowledge out of it with a small power and energy budget,” says Tomas Tuma at IBM Research-Zurich in Switzerland.

The artificial neuron is just a micrometre across, and made from a chalcogenide-based crystal sandwiched between electrodes.

Incoming information arrives as pulses of energy. This alters the temperature of the crystal and makes it change from an ordered crystalline structure to a more glass-like amorphous state. When this phase change reaches a certain stage, the crystal “fires” – emitting an electrical signal of its own, just as a neuron does. A final energy pulse then resets the crystal, returning it to its original phase.

Conventional computer chips operate as on-off switches, flipping in response to a voltage change. By instead firing only once an certain input threshold is reached, the crystal chips should be better at making sense of large amounts of chaotic data – especially when working as a pack.

Imagine you wanted to monitor thousands of Twitter accounts for tweets that mentioned IBM, says Tuma. You could have a system that notified you of every single mention. But by having the chips fire only once the number of mentions passes a given threshold, you’re more likely to pick up on something meaningful, he says.

To test the set-up, the team fired a broadband signal at packs of around 500 artificial neurons. That signal was too fast for individual chips to handle: the phase change happened but often not quickly enough to be useful, as it lagged behind the input. The result was that each chip spiked at slightly different times. But, taken together, the hundreds of chips produced a discernible pattern that represented the incoming information – a feat that resembles how groups of neurons work together in the brain.

 

Storms of data

In future applications, storms of such electronic pulses might represent information about the stock exchange, or weather events picked up by sensors around the world.

The researchers also say chips of this kind could be useful for processing data from the Square Kilometre Array, the world’s largest telescope, being built in South Africa and Australia. This will create petabytes of data every day. “Such a neuromorphic device could help filter out the garbage from what’s interesting to astronomers,” says Chris Sciacca at IBM.

Gert Cauwenberghs at the University of California, San Diego, is impressed. It’s the first demonstration of a phase-change device of this kind emulating the spiking of a biological neuron, he says.

The idea here is not to replace generic processors, says Tuma. Rather the chip could work alongside existing ones in many different types of device.

Cauwenberghs thinks it’s possible that all computers might use chips of this kind one day, especially if advantages in energy efficiency hold up in future tests. “It remains to be seen if such technology can prove its chops to the computer industry.”

IBM is working on another brain-inspired system nicknamed TrueNorth, which  also mimics the firing patterns of large groups of neurons.

DOI: Nature Nanotechnology, 10.1038/NNANO.2016.70