Task-adaptive physical reservoir computing.

Area of Science:

• Neuromorphic computing
• Materials science
• Machine learning

Background:

• Reservoir computing offers energy-efficient machine learning solutions.
• Physical reservoir computing lacks the reconfigurability of software-based methods.
• Tuning hyperparameters is crucial for adapting computing performance to tasks.

Purpose of the Study:

• To introduce a task-adaptive approach for physical reservoir computing.
• To enable reconfiguration of physical reservoir properties for diverse computational tasks.
• To overcome the limitations of fixed responses in physical reservoir computing.

Main Methods:

• Leveraging thermodynamical phase space to reconfigure reservoir properties.
• Utilizing spin-wave spectra of chiral magnets (Cu2OSeO3) with distinct magnetic phases (skyrmion, conical, helical).
• Demonstrating task-adaptability in other chiral magnets (Co8.5Zn8.5Mn3, FeGe) at room temperature.

Main Results:

• Optimized computational performance across a diverse task set using the task-adaptive approach.
• Achieved on-demand access to different computational reservoir responses by tuning magnetic phases.
• Showcased the applicability of the approach in various chiral magnets at above and near room temperatures.

Conclusions:

• The task-adaptive approach enhances the flexibility and applicability of physical reservoir computing.
• This method provides a viable solution to the energy costs associated with machine learning.
• The demonstrated room-temperature operation highlights the practical potential of this neuromorphic architecture.