Gesture recognition with Brownian reservoir computing using geometrically confined skyrmion dynamics

Area of Science:

• Physics
• Materials Science
• Computer Science

Background:

• Physical reservoir computing offers efficient information processing with reduced energy and training needs.
• Magnetic skyrmions, as topological spin textures, are ideal for reservoir computing due to stability and low-power manipulation.
• Existing spin-based reservoir computing faces limitations in quasi-static detection and real-world data temporal rescaling.

Purpose of the Study:

• To develop a time-multiplexed skyrmion reservoir computing system.
• To enable alignment of reservoir timescales with real-world temporal patterns.
• To demonstrate real-time data integration without temporal rescaling.

Main Methods:

• Utilized magnetic skyrmions as the physical reservoir.
• Implemented time-multiplexing to adjust reservoir timescales.
• Processed millisecond-scale hand gestures from Range-Doppler radar data via voltage excitations.
• Detected skyrmion trajectory evolution for information processing.

Main Results:

• Achieved competitive or superior performance compared to energy-intensive software-based neural networks.
• Demonstrated a scalable, nanometer-level hardware approach.
• Successfully integrated real-world sensor data in real-time without temporal rescaling.

Conclusions:

• Time-multiplexed skyrmion reservoir computing effectively addresses limitations of traditional methods.
• This hardware approach offers a pathway for real-time, low-power information processing.
• The technology has broad applicability in various real-time data integration scenarios.