Neural networks are one of the most important tools of artificial intelligence (AI): they mimic the functioning of the human brain and know texts, language and images reliably, to name a few. So far they work with conventional processors as adaptive software, but experts are working on an alternative concept, the “neuromorphic computer”. In this case, the points of change in the brain – the neurons – are not simulated by the software but are reconstructed in the hardware components. A team of researchers Helmholtz Center Dresden-Rossendorf (HZDR) has now shown a new approach to hardware – focused magnetic waves that are created and split in micrometer-sized leaves. Looking to the future, this means that optimization tasks and knowledge of models can be completed faster and more energy efficiently. The researchers presented the results in the journal Physical Review Letters.
The group based their research on a small disk of magnetic iron nickel material that was only a few micrometers in diameter. A golden ring is placed around this disk: when it passes through an alternating current between gigahertz, it emits microwaves that excite the so-called spin waves on the disk. “Iron nickel electrons show a kind of rotation, a kind of rotation that makes them spin,” explains Helmut Schultheiß, head of the “Magnonics” group at the Emmy Noether Group in HZDR. “We use microwave pulses to move the top of the electrons away a little bit.” Then the electrons transmit this disturbance to their neighbors – this causes a spin wave to be thrown through the material. Information can be transported very efficiently without having to move electrons themselves, which is what happens in today’s computer chips.
In 2019, the Schultheiß group found something remarkable: in some cases, the spin wave generated in the magnetic vortex can be split into two waves, each with a reduced frequency. “Non-linear effects are responsible for this,” explains Schultheiß’s colleague Lukas Körber. “Radiated microwave power is only activated when it exceeds a certain threshold.” This behavior suggests spin waves as a promising candidate for artificial neurons because of the striking parallelism with brain functioning: these neurons are only triggered when the stimulus threshold is exceeded.
Initially, however, scientists were unable to control the very precise distribution of the spin wave. Körber explains: “When we sent the microwave to the disk, there was a period of time split between the two new waves of the spin wave. And that was hard to control. ”So the team had to think of a way around the problem, as it is now described in the Physical Review Letters: In addition to the gold ring, a small magnetic stripe is attached next to the magnetic leaf. creates in this band, which can interact with the wave rotation wave, and can therefore be a kind of deception.The rotation wave in the list causes the wave wave to distribute faster. he wants us to be able to start the process now and control the time delay “.
This means that, in principle, it has been shown that spin wave leaves are suitable for artificial hardware neurons – they change like brain nerve cells and can be directly controlled. “The next thing we want to do is build a small network with our spin wave neurons,” Helmut Schultheiß announced. “This neuromorphic network should perform simple tasks, such as knowing the correct patterns.”
Optimizing facial recognition and traffic
Model knowledge is one of the main applications of AI. A face recognition on a smartphone, for example, ignores the need for a password. In order to function, a network of neurons must be trained in advance, which leads to high computational power and a large amount of data. Smartphone manufacturers transfer this network to a special chip and then integrate it into the phone. But the chip has a weakness. She’s not adaptable, so she can’t recognize the faces wearing the Covid mask, for example.
A neuromorphic computer, on the other hand, can also cope with such situations: compared to conventional chips, its components are not hard wires, they function like nerve cells in the brain. “Therefore, a neuromorphic computer can process large volumes of data at once, just like humans – and very energy efficiently,” says Schultheiß. In addition to model recognition, the new type of computer could be useful in another area that is economically important: for the optimization tasks of high-precision phone route organizers.
L. Körber, K. Schultheiss, T. Hula, R. Verba, J. Fassbender, A. Kákay, and H. Schultheiss, “Non-local stimulation of division in a magnomagnetic vortex “, November 12, 2020. Physical Review Letters.
DOI: 10.1103 / PhysRevLett.125.207203
“Excitation of Whispering Gallery Magnons in a Magnetic Vortex” by K. Schultheiss, R. Verba, F. Wehrmann, K. Wagner, L. Körber, T. Hula, T. Hache, A. Kákay, AA Awad, V. Tiberkevich, AN Slavin, J. Fassbender, and H. Schultheiss, March 5, 2019, Physical Review Letters.
DOI: 10.1103 / PhysRevLett.122.097202