Homogeneous Memristors with Tunable Decay Dynamics for Self-Adaptive Reservoir Computing

Area of Science:

• Materials Science
• Computational Neuroscience
• Electrical Engineering

Background:

• Memristors offer advantages for reservoir computing (RC) due to their nonlinear dynamics.
• Integrating different materials for reservoir nodes and readout layers presents challenges.
• Existing RC systems often have fixed nodes, limiting processing of complex temporal sequences.

Purpose of the Study:

• To develop a homogeneous RC system using CMOS-compatible materials.
• To enable tunable temporal dynamics in reservoir nodes.
• To improve the processing of complex temporal sequences and self-adaptation in RC systems.

Main Methods:

• Fabrication of a homogeneous RC system using TiO_x/AlO_y dynamic memristors for reservoir nodes and AlO_y/TiO_x for readout layers.
• Modulation of memristor time constants by controlling reading voltage bias (V_bias).
• Simulation of a self-adaptive RC system for dynamic gesture recognition.

Main Results:

• Achieved an expanded precise prediction temporal scale of 10³ in the Hénon map benchmark.
• Demonstrated improved recognition accuracy from 82.0% to 93.9% in dynamic gesture recognition.
• Successfully implemented a homogeneous and self-adaptive RC system.

Conclusions:

• The developed homogeneous RC system based on CMOS-compatible oxides enhances complex temporal sequence processing.
• Controlling memristor time constants offers an effective approach for self-adaptation and improved performance.
• This work paves the way for advanced, integrated RC systems.