This is the authors' abstract. We don't add key points for this paper.
Spiking neuron circuits, responsible for encoding analog signals into spiking signals, are crucial for conversion-based spiking neural networks (SNNs), enabling direct integration with conventional deep learning. However, the mainstream ReLU (Rectified Linear Unit) spiking neuron circuits lack the capability to encode negative values, resulting in loss of information. In this work, a spiking neuron circuit featuring the L-ReLU (Leaky Rectified Linear Unit) function based on an optimized threshold switching (TS) memristor model with an asymmetric switching character was proposed. The adjustment of the hyperparameter β in the TS memristor model enables the realization of a flexible negative coding function. Additionally, the CIFAR-10 classification task was executed by embedding the proposed L-ReLU spiking neuron circuit in the VGG16 model. The simulation results indicate that an obvious improvement was obtained for the CIFAR-10 classification task by adopting L-ReLU. This work provides a novel approach for memristor-based neuron circuits applied in conversion-based SNN.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zou et al. (2024) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: