A Self-Calibratable Spiking Neuron based on Dual-Mode Memritor for High-Accuracy Spiking Neural Networks

In this study, a high-speed and self-calibratable neuromorphic neuron is demonstrated based on dual-mode memristor devices with dual switching modes. By tuning the resistance of the RRAM, the device-to-device (D2D) variation between different neurons can be substantially regulated, thereby significantly improving the uniformity of neural spiking frequencies of different neurons under the same stimulus. For a three-layer fully connected network designed for MNIST dataset, the accuracy of the network increased from 89.70% to 90.03%. This research addresses the D2D issue among spiking neurons and advances the development of highly reliable neuromorphic computing.