Highly Reliable Bi₂O₂Se Dendritic Neuron Enabling Spatial-Temporal Signal Processing for Real-World Image...

Area of Science:

• Materials Science
• Neuroscience
• Artificial Intelligence

Background:

• Current artificial intelligence (AI) models simplify biological neurons, limiting performance and energy efficiency in complex tasks.
• Biological neurons process information nonlinearly via dendrites, capturing spatial-temporal data crucial for advanced computation.

Purpose of the Study:

• To demonstrate a compact artificial dendrite device utilizing memristors based on bismuth oxyselenide (Bi₂O₂Se).
• To implement a dendritic neuron model and evaluate its performance in a neural network for pattern recognition.

Main Methods:

• Direct growth of transfer-free Bi₂O₂Se switching medium on metal-patterned substrates using a 350 °C selenization process.
• Fabrication of memristive devices leveraging the layered structure of Bi₂O₂Se to limit metal injection and ensure stable switching.
• Modeling the device's current response to spatial-temporal voltage inputs and implementing a dendritic neuron model.

Main Results:

• Achieved reliable dynamic resistive switching with excellent cycle uniformity and over 2 million cycles of exceptional endurance.
• Demonstrated a highly reliable current response characteristic of the Bi₂O₂Se memristor.
• Implemented neural network using the Bi₂O₂Se dendrite device achieved a 78.3% recognition rate on the Street View House Numbers (SVHN) dataset.

Conclusions:

• The Bi₂O₂Se-based artificial dendrite device effectively mimics biological neuron functionalities, offering a pathway to more efficient and powerful AI.
• The developed memristor technology shows significant potential for next-generation neuromorphic computing and advanced AI applications.