Volatile self-selective memristive neuron for millisecond-latency neuromorphic object detection at the edge

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence Hardware
• Materials Science

Background:

• Conventional computing architectures face power and latency constraints for real-time AI object detection in edge applications.
• Developing efficient hardware is crucial for advancing edge AI capabilities.

Purpose of the Study:

• To present a bio-inspired neuromorphic system utilizing self-selective GaOx/ZnO memristors.
• To address the limitations of conventional computing for edge AI object detection.

Main Methods:

• Fabrication and characterization of GaOx/ZnO memristors with high selectivity and nonlinearity.
• Integration of memristors into a 32x32 array emulating the frog visual system's first-spike-time-coding.
• Implementation of the system for aerial drone object detection.

Main Results:

• The memristor device demonstrated a selection ratio and nonlinearity of ~10⁷, low leakage currents, and microsecond-scale volatile dynamics.
• The neuromorphic array achieved millisecond-scale pulse responses and reliable object recognition for pedestrians and vehicles with a minimal accuracy drop (2.5%) compared to simulations.
• The array exhibited a parallel processing capability of ~8.36×10¹² computational nodes.

Conclusions:

• This work offers a practical hardware solution for edge neuromorphic computing systems.
• The developed system enables fast-response object detection suitable for intelligent transportation and real-time monitoring applications.