Credit: CC0 Public Domain
Researchers have demonstrated that artificial intelligence may be performed using small nanomagnets that interact similarly to neurons in the brain.
The new technology, created by a team lead by Imperial College London researchers, has the potential to significantly reduce the energy costs associated with artificial intelligence (AI), which are now doubling every 3.5 months on a worldwide scale.
The worldwide team demonstrated for the first time in an article published today in Nature Nanotechnology that networks of nanomagnets may be employed to perform AI-like processing. Nanomagnets may be utilised for ‘time-series prediction’ activities, such as forecasting and managing insulin levels in diabetic patients, the researchers demonstrated.
Artificial intelligence based on ‘neural networks’ tries to recreate the way neurons communicate with one another to process and store information in the brain. Much of the mathematics that underpins neural networks was established by physicists to describe the way magnets interact, but it was too difficult to employ magnets directly at the time due to researchers’ lack of understanding of how to put data in and get information out.
Rather than that, software running on conventional silicon-based computers was utilised to recreate the magnet interactions, recreating the brain in the process. Now, the team has been able to interpret and store data directly on the magnets, eliminating the software simulation’s middleman and potentially resulting in substantial energy savings.
States of nanomagnetism
Nanomagnets can exist in a variety of different’states’ depending on their orientation. When a magnetic field is applied to a network of nanomagnets, the state of the magnets varies depending on the parameters of the input field, but also on the surrounding magnets’ states.
The team, led by Imperial Department of Physics experts, then devised a method for counting the number of magnets in each condition after the field passed through, providing the ‘answer’.
Dr. Jack Gartside, co-first author of the work, stated: “For a long time, we’ve been attempting to solve the challenge of how to enter data, pose a question, and receive a response using magnetic computing. Now that we have demonstrated that it is possible, it paves the way for the abolition of the computer software that does the energy-intensive simulation.”
“How the magnets interact gives us all the information we need; the laws of physics themselves become the computer,” co-first author Kilian Stenning remarked.
Dr. Will Branford, the team’s leader, stated: “It has been a long-term objective to implement computer hardware inspired by Sherrington and Kirkpatrick’s software algorithms. It was not possible to provide the requisite control and readout using the spins on atoms in ordinary magnets, but by scaling up the spins into nanopatterned arrays, we were able to.”
Cost reductions in energy
Artificial intelligence is already being employed in a variety of scenarios, ranging from voice recognition to self-driving cars. However, training AI to perform even relatively simple tasks can use a significant amount of energy. For instance, training artificial intelligence to solve a Rubik’s cube consumed the energy equivalent of two nuclear power plants operating for an hour.
Much of the energy consumed by typical silicon-chip computers to accomplish this is squandered on inefficient electron transport during processing and memory storage. Nanomagnets, on the other hand, do not rely on the physical passage of particles such as electrons, but rather process and transfer information via a’magnon’ wave, in which each magnet has an effect on the state of surrounding magnets.
This results in significant energy savings and enables the processing and storage of information to occur concurrently, rather than as distinct processes as in conventional computers. This breakthrough could increase the efficiency of nanomagnetic computing by up to 100,000 times compared to conventional computing.
At the frontier of artificial intelligence
The scientists will then train the system using real-world data, such as ECG readings, with the goal of eventually turning it into a functioning computer device. Magnetic systems may eventually be integrated into conventional computers to increase the energy efficiency of high-performance computing activities.
Their energy efficiency also enables them to be fueled by renewable energy and used for ‘AI at the edge’—processing data on-site, such as at weather stations in Antarctica, rather than transferring it to massive data centres.
Additionally, they could be employed on wearable devices to interpret biometric data from the body, such as predicting and managing insulin levels in diabetic patients or detecting irregular heartbeats.
Further information: Jack Gartside, Reconfigurable training and reservoir computing in an artificial spin-vortex ice via spin-wave fingerprinting, Nature Nanotechnology (2022). DOI: 10.1038/s41565-022-01091-7. www.nature.com/articles/s41565-022-01091-7
Journal information: Nature Nanotechnology
Source: Imperial College London