New research artificially creating a rare form of matter known as spin glass
could spark a new paradigm in artificial intelligence by allowing algorithms
to be directly printed as physical hardware. The unusual properties of spin
glass enable a form of AI that can recognize objects from partial images
much like the brain does and show promise for low-power computing, among
other intriguing capabilities.
"Our work accomplished the first experimental realization of an artificial
spin glass consisting of nanomagnets arranged to replicate a neural
network," said Michael Saccone, a post-doctoral researcher in theoretical
physics at Los Alamos National Laboratory and lead author of the new paper
in Nature Physics. "Our paper lays the groundwork we need to use these
physical systems practically."
Spin glasses are a way to think about material structure mathematically.
Being free, for the first time, to tweak the interaction within these
systems using electron-beam lithography makes it possible to represent a
variety of computing problems in spin-glass networks, Saccone said.
At the intersection of engineered materials and computation, spin-glass
systems are a type of disordered system of nanomagnets arising from random
interactions and competition between two types of magnetic order in the
material. They exhibit "frustration," meaning that they don't settle into a
uniformly ordered configuration when their temperature drops, and they
possess distinct thermodynamic and dynamic traits that can be harnessed for
computing applications.
"Theoretical models describing spin glasses are broadly used in other
complex systems, such as those describing brain function, error-correcting
codes or stock-market dynamics," Saccone said. "This wide interest in spin
glasses provides strong motivation to generate an artificial spin glass."
The research team combined theoretical and experimental work to fabricate
and observe the artificial spin glass as a proof-of-principle Hopfield
neural network, which mathematically models associative memory to guide the
disorder of the artificial spin systems.
Spin glass and Hopfield networks have developed symbiotically, one field
feeding off the other. Associative memory, whether in a Hopfield network or
other forms of neural networks, links two or more memory patterns related to
an object. If just one memory is triggered—for instance, by receiving a
partial image of a face as input—then the network can recall the complete
face. Unlike more traditional algorithms, associative memory does not
require a perfectly identical scenario to identify a memory.
The memories of these networks correspond to ground states of a spin system
and are less disturbed by noise than other neural networks.
The research by Saccone and the team confirmed that the material was a spin
glass, evidence that will allow them to describe the properties of the
system and how it processes information. AI algorithms developed in spin
glass would be "messier" than traditional algorithms, Saccone said, but also
more flexible for some AI applications.
Reference:
Michael Saccone et al, Direct observation of a dynamical glass transition in
a nanomagnetic artificial Hopfield network, Nature Physics (2022).
DOI: 10.1038/s41567-022-01538-7