A phase-change material held between two electrodes doesn’t switch completely between ordered and jumbled arrangements like ones and zeros. Instead, at any point in time, there is a mix of both: the overall resistance of the material is determined by the size of the regions where atoms are jumbled.

“We’re coding information in terms of atomic arrangements,” says Sebastian. The weights of a neural network, for example, can be stored and accessed as the resistance in a phase-change memory device.

But these resistances suffer from drift and fluctuation. Because current passes through the phase-change material when information is read, the jumbled regions change a little bit every time—which has limited the precision and practicality of such devices.

To circumvent this problem, the IBM researchers introduced a so-called ‘projection segment’ to the phase-change memory device. First proposed in 2015 by the same team, the projection segment is a conducting layer of metal nitride that wraps around a phase-change material core and runs parallel to it between electrodes. The projection segment separates the information writing and reading processes.

This projection segment does nothing when information is being written; all the current runs through the phase-change material to tweak the jumbled regions. But when information is being retrieved, the current flows through the projection segment and around the jumbled regions, leaving them untouched and preserving the information that’s stored. “That is the key innovation here,” says Sebastian.

The researchers tested a single-layer neural net on an 8-bit chip composed of 30 phase-change memory devices to identify pictures of the digits 1, 0, and 4, and achieved 100 percent classification accuracy. While premature, Sebastian estimates the advance could potentially bring some 100 to 1,000-fold gains in power savings to future devices, compared to traditional computing.

Precision was sought in traditional computing, but with artificial intelligence, there is now an opposite trend. IBM is also reporting a digital chip today that operates at 8-bits while maintaining accuracy in neural net training. That models more closely to the human brain, which often can draw correct conclusions from little information.

IBM’s vice-president for research Jeff Welser likens this to looking out a foggy window and seeing a blurry person walking toward your house. “As soon as you recognize your mom, it doesn’t matter how low-precision the image is,” says Welser. “You’ve got the right information you need.”