Developing New Hardware That Meets the Energy Demands of AI Applications

To try and solve this problem, researchers at Purdue University are developing hardware that is able to learn skills using a type of AI that currently runs on and is typically reserved for software platforms. The researchers think that this approach could offset the energy needed for using AI in certain advanced applications like autonomous vehicles because intelligence features would be shared between hardware and software. 

“Software is taking on most of the challenges in AI. 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,” said Shriram Ramanathan, a professor of materials engineering at Purdue University. 

In their research study, the researchers describe how applying “tree-like” memory features in spiking neural networks to demonstrate high fidelity object recognition could open up new directions for artificial intelligence. 

 

Demonstrating ‘Tree-like’ Memory in Hardware

The software uses tree-like memory to sort and organize information into “branches” so that it can be retrieved efficiently, a strategy inspired by the human brain. “Humans memorize things in a tree structure of categories,” said Hai-Tian Zhang, a Purdue postdoctoral fellow. "For example, we memorize apples and oranges under ‘fruit’ and cats and dogs under ‘animal’. “Mimicking these features in hardware is potentially interesting for brain-inspired computing,” he added. 

In their new study, the research scientists described the development of a new type of hardware component made from a quantum material, neodymium nickel oxide, that exhibited an artificial tree-like memory. Although tree-like memory has been demonstrated in potential hardware before, the Purdue team managed to observe the tree-like memory at room temperature whereas previous works have only observed it in hardware at temperatures too low for electronic devices.

 

The new hardware for artificial intelligence developed by Purdue University researchers. Image credited to Purdue University

 

Taking Advantage of ‘Quantum Mechanical Effects’ 

The team introduced a proton to the so-called quantum material which would move when an electrical pulse was applied. Quantum materials are known for having properties that cannot be explained by classical physics. 

Each time that the proton moved, a new resistance state was observed which created an information storage site or ‘memory states’. The team discovered that several of these pulses could be used to create a branch made up of memory states and that thousands of memory states in the material can be built to take advantage of quantum mechanical effects. Through simulations, the team showed that the material can learn the numbers 0 to nine, a baseline test of artificial intelligence. 

 

Sharing Intelligence Features

Although AI hardware is still in its infancy, researchers have already demonstrated AI in pieces of potential hardware. However, its large energy demand has not yet been addressed. The Purdue research team believes that their demonstration of this tree-like structure at room temperature in a material is a step in the right direction and shows that hardware can offload tasks from the software. 

"This discovery opens up new frontiers for AI that have been largely ignored because implementing this kind of intelligence into electronic hardware didn't exist," Ramanathan said.