Energy-efficient memristor-based spiking neural network for edge devices with a novel window function

Abstract

The conventional artificial neural network is not suitable for the development trend of edge artificial intelligence due to its high computational energy requirements. In this study, we propose an energy-efficient system using spiking neural networks based on a memristor crossbar. A novel window function is introduced, which overcomes the shortcomings of conventional window functions. Additionally, a dynamic learning rate matrix approach is suggested to decrease the influence of conductance drift and conductance noise on neural networks, efficiently eliminate noise, and adjust the learning rate for each individual synapse. We evaluate the performance of the proposed method using an energy consumption evaluation model. Experimental results show that the proposed window function outperforms state-of-the-art window functions in terms of accuracy and test time. Furthermore, the dynamic learning rate matrix algorithm achieves 97.27% accuracy on the MNIST dataset. Memristor-based spiking neural networks have a significant energy consumption advantage over conventional artificial neural networks, making this approach suitable for resource-constrained edge artificial intelligence devices.