Feature-Selective Preprocessing with Electrically Robust Boron Nitride-Based Dynamic Memristors for Reliable...

Area of Science:

• Materials Science
• Computer Science
• Artificial Intelligence

Background:

• Reservoir computing (RC) offers efficient AI computation via dynamic data preprocessing.
• Existing RC methods face accuracy limitations due to data loss and physical reservoir switching endurance issues.

Purpose of the Study:

• To develop an electrically robust memristor for dynamic reservoir computing applications.
• To address accuracy limitations and switching endurance challenges in current RC systems.

Main Methods:

• Fabrication of a Cu/a-BN/Ti(TiO X) dynamic memristor with a built-in series resistance structure.
• Characterization of the memristor's volatile, gradual switching, and reservoir dynamics.
• Simulation of noisy image classification using the developed memristor for RC and Wide RC (WRC) with Convolutional Neural Networks (CNNs).

Main Results:

• The memristor exhibited excellent endurance (1.2 million switching updates) and uniformity, suppressing filament overgrowth.
• RC implementation simplified image processing by selecting key spatial features, enhancing network efficiency.
• The proposed WRC/CNN architecture achieved high accuracy with minimal energy increase, enabling lightweight edge AI.

Conclusions:

• The novel memristor provides a robust and efficient solution for RC, overcoming previous limitations.
• The developed WRC/CNN architecture is suitable for high-accuracy, low-power AI on edge devices.