Researchers Create Perovskite "Tree" Memories for More Natural, Energy-Efficient AI Hardware

Researchers at Purdue University, the Universities of Illinois, Louisville, Iowa, and California San Diego claim to have designed hardware which can replace power-hungry artificial intelligence software running on general-purpose platforms — and in doing so dramatically reduce its power usage.

"Software is taking on most of the challenges in AI," claims Purdue's Professor Shiram Ramathan, pointing to an article published in Wired earlier this year claiming that an artificial intelligence algorithm for solving a Rubik's cube could chew through the output of three nuclear power plants — all by itself. "If you could incorporate intelligence into the circuit components in addition to what is happening in software, you could do things that simply cannot be done today."

"Humans memorise things in a tree structure of categories. We memorise 'apple' under the category of 'fruit' and 'elephant' under the category of 'animal,' for example," explains Hai-Tian Zhang, the paper's lead author. "Mimicking these features in hardware is potentially interesting for brain-inspired computing."

Zhang and team have developed a quantum material based on perovskites, a material being investigated for boosting the efficiency of solar cells. In the team's hands, though, that same material has been used to produce tree-like memory cells which are capable of operating at room temperature — a major breakthrough over previous attempts, which have required chilling to extremely low temperatures.

The key to the memory is the introduction of a proton to neodymium nickel oxide; applying an electrical pulse moves the proton and adjusts the resistance of the material, allowing for the creation of a memory state — and multiple pulses create multiple memory states.

"We can build up many thousands of memory states in the material by taking advantage of quantum mechanical effects," says Ramanathan. "The material stays the same. We are simply shuffling around protons."

In simulation, the material was able to "learn" the numbers 0 through 9 - a baseline test for artificial intelligence. "Protons also are natural information transporters in human beings," adds Zhang of the material's potential. "A device enabled by proton transport may be a key component for eventually achieving direct communication with organisms, such as through a brain implant."

The team's work has been published in the journal Nature Communications under open-access terms.