EWSbp-SNN: Error-Weighted STDP With Backpropagation for Data Classification in Spiking Neural Networks

Spiking neural networks (SNNs) have exhibited remarkable potential in neuromorphic data classification, especially in processing dynamic vision sensor (DVS) data. However, SNNs still have challenges in improving classification accuracy and in mitigating vanishing and exploding gradients. To address these challenges, we propose the error-weighted STDP with backpropagation for data classification in spiking neural networks (EWSbp-SNN). It combines error-weighted STDP (EWS) with backpropagation (BP) for weight updates, and incorporates dynamic threshold batch normalization (DTBN). Firstly, in order to improve the classification accuracy, the local EWS approach utilizes a weight factor calculated based on the network prediction error. This factor integrates the importance information of connections from the previous layer to perform the first local weight update, effectively suppressing the influence of irrelevant or noisy features. Afterward, BP is applied to perform the second global weight update, ensuring consistent optimization across time steps. In addition, to address the problem of vanishing and exploding gradients, the DTBN technique is presented, which normalizes data based on dynamic threshold variations during forward propagation, ensuring timely gradient adjustments in the local weight updates. Experiment results demonstrate that the proposed method significantly improves training accuracy and achieves state-of-the-art performance in gradient smoothing for classification tasks. Especially, it achieves an average accuracy of 8.02% higher than other methods on the DVS-CIFAR10 dataset. The source code will be available at https://github.com/linglingMiu/EWSNN.