A Learning-Rate Modulable and Reliable TiO_x Memristor Array for Robust, Fast, and Accurate Neuromorphic Computing.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Neuromorphic hardware systems are crucial for energy-efficient big data processing and AI.
• Memristor-based systems offer a promising approach for hardware neural networks.
• Titanium oxide (TiO_x) memristors are key components for these systems.

Purpose of the Study:

• To fabricate uniform and reliable TiO_x memristor array devices.
• To design and implement an in situ convolutional neural network hardware system using these memristors.
• To demonstrate efficient in situ training with learning rate modulation.

Main Methods:

• Fabrication of multiple 25 × 25 TiO_x memristor arrays.
• Development of a global constant voltage programming scheme for crossbar arrays.
• Implementation of in situ training with direct memristor operation and learning rate modulation.

Main Results:

• Achieved high device uniformity (threshold uniformity ≈2.7%) and yield (>99%).
• Demonstrated superior memristor performance including repetitive stability (≈3000 spikes) and ambient stability (6 months).
• Enabled fast-converging in situ training with five times fewer iterations and reduced energy consumption, achieving ≈95.2% classification accuracy.

Conclusions:

• Uniform TiO_x memristor arrays are suitable for energy-efficient neuromorphic hardware.
• The developed in situ training method significantly reduces training iterations and energy.
• This approach advances the realization of practical AI hardware systems.