Adjustable short-term memory of SiO_x:Ag-based memristor for reservoir computing

Area of Science:

• Materials Science and Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Temporal information processing is vital across finance, biomedicine, and engineering.
• Reservoir computing (RC) offers efficient temporal data processing with low training costs.
• Existing memristor-based RC systems have fixed short-term memory, limiting task adaptability.

Purpose of the Study:

• To propose and demonstrate methods for modulating the short-term memory of Pt/SiO:Ag/Pt memristors.
• To enhance the performance of reservoir computing systems by adjusting memristor characteristics.
• To enable memristor-based RC systems for diverse temporal analysis tasks.

Main Methods:

• Modulating memristor short-term memory by controlling read voltage, pulse amplitude, and pulse width.
• Fabricating Pt/SiO:Ag/Pt memristors with adjustable characteristic times (microseconds to milliseconds).
• Implementing memristor-based RC networks for classification, time-series prediction, and pattern recognition.

Main Results:

• Achieved a three-order-of-magnitude range in characteristic time for memristors.
• Demonstrated high accuracy in 4-bit pulse stream classification (99.6% spoken digits, 91.7% MNIST).
• Obtained low prediction errors for Hénon map (0.003) and Mackey-Glass time series (0.27).

Conclusions:

• Memristor-based RC systems with adjustable short-term memory show superior performance.
• The developed approach offers high controllability for handling multiple temporal processing tasks.
• This work advances the application of memristors in adaptive and efficient temporal information processing.