Quantum materials known as Mott insulators can “learn” to respond to external stimuli in a way that mimics animal behaviour, say researchers at Rutgers University in the US. The discovery of behaviours such as habituation and sensitization in these non-living systems could lead to new algorithms for artificial intelligence (AI).
Neuromorphic, or brain-inspired, computers aim to mimic the neural systems of living species at the physical level of neurons (brain nerve cells) and synapses (the connections between neurons). Each of the 100 billion neurons in the human brain, for example, receives electrical inputs from some of its neighbours and then “fires” an electrical output to others when the sum of the inputs exceeds a certain threshold. This process, also known as “spiking”, can be reproduced in nanoscale devices such as spintronic oscillators. As well as being potentially much faster and energy efficient than conventional computers, devices based on these neuromorphic principles might be able to learn how to perform new tasks without being directly programmed to accomplish them.
Mimicking non-associative learning
In the new work, researchers led by Subhasish Mandal in the Department of Physics and Astronomy at Rutgers focused on nickel oxide, which is a typical example of a Mott insulating material. When they monitored how the material’s electrical conductivity changes as the concentration of its atomic defects is reversibly modulated using external stimuli such as oxygen, ozone and light, they found that it mimics non-associative learning. This type of learning is one of the most fundamental ways in which living organisms learn, and it helps animals adapt continuously to changing situations.
Another intriguing finding from the study is that when the researchers exposed the nickel oxide to rapidly changing oxygen concentrations or different light intensities, the material was unable to respond fully, and instead remained in an unstable state with its electrical conductivity fluctuating little. When they later introduced additional atomic defects into the sample using a harsher stimulus, ozone, the material’s electrical conductivity fluctuated faster, only to slow down again.
Universal learning characteristics
Mandal says that the team’s results demonstrate universal learning characteristics such as habituation and sensitization that are generally found in living species. He also suggests that the characteristics they unearthed could inspire new algorithms for unsupervised learning in neural networks and AI, much as the collective motion of birds or fish has done in the past. “The growing field of AI requires hardware that can host adaptive memory properties beyond what is used in today’s computers,” he says. “We find that nickel oxide insulators, which historically have been restricted to academic pursuits, might be interesting candidates to be tested in the future for brain-inspired computers and robotics.”
AI and particle physics: a powerful partnership
Beyond nickel oxide, the researchers say that similar effects could be found in other correlated materials that have defects susceptible to modulation using external stimuli. They now plan to further explore the learning behaviour of nickel oxide devices under electric fields in test chips.
The team reports its work in PNAS.