A Self-Organizing Map Spiking Neural Network Based on Tin Oxide Memristive Synapses and Neurons

Abstract

Neuromorphic computing systems are promising alternatives in areas such as pattern recognition and image processing. This work focuses on the fabrication of tin oxide memristors (Ag/SnO₂/Pt) to emulate artificial synapses and neurons. These tin oxide memristors demonstrate stable switching between threshold switch (TS) and resistive switch (RS) modes, achieved by adjusting the compliance current. Notably, this memristor achieves extremely low switching voltage and excellent cycle endurance. Moreover, the conductance value of the memristor can continuously transform under different illumination conditions, such as white light and purple light. A single tin oxide memristor device is used to model typical neuromorphic responses, such as synaptic plasticity and artificial neuron impulse responses. This approach offers a promising solution for high-density, low-power, brain-inspired computing chips. Additionally, memristive Leaky Integrate-and-Fire (LIF) neuron and synapse models are designed and integrated for the first time into a Self-Organizing Map Spiking Neural Network (SOM-SNN) architecture. Applying this architecture to an unsupervised learning self-organizing map memristor SNN achieved an impressive 94% recognition rate on the MNIST dataset. This study elucidates the potential for seamlessly integrating memristors into neuromorphic systems.