Spiking Neural Network with Silicon Oxide Memristive Devices in the Subthreshold Regime

This study explores silicon oxide-based memristive devices in spiking neural networks (SNNs), leveraging the subthreshold regime for performance and robustness. A bio-inspired homeostasis-driven dropout mechanism dynamically adjusts synaptic conductance to counteract memristor non-idealities like variability and noise. This adaptive regularization stabilizes networks and reduces overfitting. Experimental results show that convolutional SNNs using this approach achieve superior accuracy, outperforming models with faulty devices. These findings highlight the potential of subthreshold memristive SNNs for scalable, low-power neuromorphic computing, with future applications in bio-signal processing and anomaly detection.