Researchers have successfully improved the approach of Brownian reservoir computing by capturing and transmitting hand gestures to the system, which then used skyrmions to identify these gestures.
At Johannes Gutenberg University Mainz (JGU), researchers have refined the Brownian reservoir computing framework by recording and conveying hand gestures to the system that utilizes skyrmions to distinguish these gestures. “We were thrilled to discover that our hardware method and concept performed remarkably well—surpassing even the energy-heavy software solutions that rely on neural networks,” stated Grischa Beneke, a member of Professor Mathias Kläui’s team at the JGU Institute of Physics. Working alongside other experimental and theoretical physicists, Beneke showcased that basic hand gestures can be accurately recognized using Brownian reservoir computing.
Reservoir computing eliminates training needs and cuts energy use
Reservoir computing systems resemble artificial neural networks but have the advantage of not requiring extensive training, leading to lower energy consumption. “We only need to train a straightforward output mechanism to translate the results,” Beneke explained. The specifics of the computing processes remain unclear and are not crucial. You can think of the system as a pond with stones tossed into it, creating a complex pattern of waves. Just as the waves indicate the number and position of the thrown stones, the system’s output provides information about the initial input.
In a recent paper published in Nature Communications, the researchers outline how they recorded simple gestures like swiping left or right using Range-Doppler radar, utilizing two radar sensors from Infineon Technologies. This radar data is transformed into voltages that are fed into a multilayered thin film forming a triangular shape with contacts at each corner. Two contacts supply the voltage that makes the skyrmion move within the triangle. “In response to the applied signals, we detect intricate motions,” explained Grischa Beneke. He added, “These skyrmion movements allow us to infer the actions that the radar system has detected.” Skyrmions are helical magnetic structures that possess significant potential for unconventional computing and as data carriers in new storage technologies. Professor Mathias Kläui, the research lead at JGU, pointed out, “Skyrmions are truly remarkable. Initially, we viewed them only as candidates for data storage, but they also present great application opportunities in computing when combined with sensor systems.”
By comparing the outcomes from Brownian reservoir computing to those from a software-based system, the researchers found that gesture recognition accuracy is comparable, if not superior, when using the Brownian approach. The integration of reservoir computing with a Brownian computing model allows skyrmions to move randomly, as local magnetic property variations have a minimal influence on their behavior. This means skyrmions can be maneuvered with very low current, presenting a notable enhancement in energy efficiency compared to software methods. As the data from the Doppler radar and the intrinsic dynamics of the reservoir operate on similar timeframes, the sensor data can be directly input into the reservoir. The timeframe of the system can be adjusted to tackle various problems.
“We find that our hardware reservoir detects radar data for different hand gestures with a fidelity at least as good as advanced software-based neural network methods,” the researchers concluded in their paper in Nature Communications. Beneke noted that there is potential for further refinement of the read-out process, which currently uses a magneto-optical Kerr-effect (MOKE) microscope. Adopting a magnetic tunnel junction might help in minimizing the system’s size. The signals generated by a magnetic tunnel junction are already being emulated to showcase the reservoir’s capability.
