A temporally coded multilayer spiking neural network and its memristor-based hardware implementation

Artificial neural networks (ANNs) have demonstrated remarkable progress in various domains. However, ANNs suffer from enormous time and energy consumption during training and inference processes. Brain-inspired spiking neural networks (SNNs) have recently attracted more attention due to their higher biological plausibility and potential cost-efficient properties. However, most existing SNNs significantly degrade in performance and efficiency when simulated on conventional CPU/GPU hardware. Therefore, a novel temporally coded multilayer SNN (TMSNN) is proposed in this study. It is a typical event-driven model, which encodes information in the relative timing of spikes rather than in firing rates and uses the leaky integrate-and-fire neuron as the basic unit to pursue high biological plausibility. Its multilayer architecture enables the model to solve complicated problems effectively. On the other hand, the proposed TMSNN can be implemented on memristor-based hardware, which uses customized weight quantization and sharing techniques to mitigate the size restrictions of the memristor crossbars. After refining the weights using the simulated annealing algorithm, the hardware implementation of TMSNN can achieve very competitive performance on benchmark datasets, outperforming state-of-the-art temporally coded SNNs in our experiments. The source code of TMSNN is available at https://github.com/jhc050998/Memristor-Crossbar-Based-SNN. © 2025 Elsevier B.V.