Latest News

sciencenews.png

RIKEN and the University of Tokyo design a skyrmion-based artificial intelligence device that can recognize handwritten numbers

2022.11.09

An international research group led by visiting scientist Tomoyuki Yokouchi of the RIKEN Center for Emergent Matter Science (and Research associate of the Graduate School of Arts and Sciences, the University of Tokyo) and Professor Yoshichika Otani of the Institute for Solid State Physics, the University of Tokyo (and Team Leader at the RIKEN Center for Emergent Matter Science), has discovered that the deformation of skyrmions induced by a magnetic field can be applied to a physical reservoir, a type of artificial intelligence device, and designed a skyrmion-based artificial intelligence device that could actually recognize handwritten numbers. They also investigated the relationship between the number of skyrmions and recognition accuracy and found that the recognition rate improves as the number of skyrmions increases. The group's findings were published in Science Advances.

Artificial intelligence is increasingly being used in a variety of areas, but the problem is that the miniaturization of devices is nearing its limits, and they consume a lot of power when executed with conventional computing elements.

In recent years, there has been active research on artificial intelligence devices (neuromorphic devices) specifically designed to execute artificial intelligence. One of these is the use of skyrmions (particle-like, topological spin structures). Skyrmions can be operated with low power consumption, and because they are nano-sized, they are expected to bring about highly integrated and high-performance artificial intelligence devices with low power consumption.

Skyrmions change in size and are created or lost when a magnetic field is applied. The research group confirmed that the deformation of the skyrmions by the magnetic field meets the properties required for a physical reservoir device, a type of artificial intelligence device. A physical reservoir device is required to convert the input signal into a nonlinear output, and the output must depend not only on the current input but also on past inputs.

First, the research group deposited and processed a multilayer thin film of platinum, cobalt and iridium, from which the skyrmions are formed, into a crisscross shape. They connected these cross-shaped elements in parallel to fabricate a skyrmion physical reservoir device.

In this skyrmion physical reservoir device, the input is the magnetic field, and the output is the value of the magnetization. The value of magnetization then reflects the state of the skyrmion. The research group then examined in detail the state of the skyrmion and the value of the magnetization when an AC magnetic field was applied using Kerr microscopy (to observe the magnetic structure) and the anomalous Hall effect (which reflects the value of the magnetization), respectively. The results show that the deformation and magnetization values of the skyrmions are nonlinear with respect to the input magnetic field and depend not only on the current input but also on past input signals.

In other words, the magnetic-field-induced skyrmion deformation has the properties required for physical reservoir devices, indicating that they can be used for this type of device.

Next, they confirmed that the skyrmion physical reservoir device can be used to perform a waveform recognition task, a performance evaluation method for artificial intelligence devices. In addition, they fabricated several physical reservoir devices with different numbers of formed skyrmions, and examined the recognition rate in each waveform recognition task. They found that the recognition rate of waveforms tends to increase as the number of skyrmions increases. This indicates that the use of skyrmions may lead to higher-performance physical reservoir devices.

They also confirmed that the skyrmion physical reservoir device can be used to perform the handwritten digit recognition task, a more complex and practical identification problem than waveform identification. First, handwritten numbers from zero to nine written by various people were preprocessed so that they could be entered into the skyrmion physical reservoir device. A total of 13,000 handwritten number data were entered into the skyrmion physical reservoir device to train it. Then, when the group tested to see if the device could identify handwritten number data from zero to nine that were not included in the training data, the skyrmion physical reservoir device obtained a recognition rate of nearly 95%. This recognition rate is comparable to the recognition rate for handwritten digit identification in previously reported artificial intelligence devices, indicating that skyrmions have the potential to be used as an artificial intelligence device.

This study found that it is possible to fabricate artificial intelligence devices using the deformation of skyrmions. However, the magnetic field used as the input signal means that it is difficult to miniaturize the device. Moving forward, further research on skyrmion-based artificial intelligence devices that use signals suitable for miniaturization, such as electric currents and surface acoustic waves as inputs instead of magnetic fields is expected to lead to the realization of low-power, highly integrated, and high-performance artificial intelligence devices using skyrmions.

Journal Information
Publication: Science Advances
Title: Pattern recognition with neuromorphic computing using magnetic-field induced dynamics of skyrmions
DOI: 10.1126/sciadv.abq5652.

This article has been translated by JST with permission from The Science News Ltd.(https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.

Back to Latest News

Latest News

Recent Updates

Most Viewed