Nanomaterials are increasingly being used for a range of applications, but their use for artificial intelligence is complicated due to the unpredictability of their behavior.
Researchers from Washington University in St. Louis believe they've come up with a model that better explains how electrons move through nanomaterials, which could, in turn, prove a significant step towards using such materials within a machine learning device.
"When one builds devices out of nanomaterials, they don't always behave like they would for a bulk material," the authors say. "One of the things that change dramatically is the way in which these electrons move through material called the electron transport mechanism, but it's not well understood how that happens."
The team developed a model in which each nanoparticle was a node in a network, and each node was connected to every other node. The model produces observable current hotspots at the nanoscale. This was further developed with a 2nd model that was based around neural networks.
"If we have a huge number of nodes — much larger than anything that exists — and a huge number of connections, how do we train it?" the researchers ask. "We want to get this large network to perform something useful, such as a pattern-recognition task."
Buoyed by the initial success, the team propose to build a simple chip with these properties and monitor its performance.
"If we treat it as a neural network, we want to see if the output from the device will depend on the input," they say. "Once we can prove that, we'll take the next step and propose a new device that allows us to train this system to perform a simple pattern-recognition task."