河北科技大学学报2025，Vol.46Issue(5)：508-520,13.DOI:10.7535/hbkd.2025yx05004
AR-SNN:脉冲神经网络鲁棒性研究
AR-SNN:Research on the robustness of spiking neural network
摘要
Abstract
Aiming at the problem of decreased robustness of spiking neural network(SNN)caused by multiple factors,an adaptive robust spiking neural network(AR-SNN)model was proposed.The model included three modules:spiking-gated linear units(S-GLU),adaptive-topK loss(A-TopK Loss),and spiking-multilayer perceptron(S-MLP).Firstly,the gating mechanism was introduced as the preprocessing layer.By improving the gated linear unit(GLU),the number of linear layers was reduced,and the S-GLU module was constructed.Secondly,the A-TopK Loss was proposed.The average loss of the samples corresponding to the top 90%of the total loss was calculated based on the proportion of cumulative losses as the final loss.Thirdly,a self-supervised learning strategy was adopted,with the multilayer perceptron(MLP)as the decoding layer to construct the S-MLP denoising network and reconstruct the original data.Finally,the experiment was conducted on the SHD speech dataset.The results show that the S-GLU module enhances the model's attention to key information and reduces the occurrence of misclassification.The A-TopK Loss enables the model to automatically focus on samples with large losses,improving its learning ability on complex data.S-MLP enhances the feature extraction ability of the network and demonstrates certain robustness to input disturbances in noise tests.The performance of the AR-SNN model is superior to that of the original model and other SNN models,and it can effectively improve the robustness of SNN.关键词
计算机神经网络/脉冲神经网络/鲁棒性/门控机制/损失函数/多层感知机Key words
computer neural network/spiking neural network/robustness/gating mechanism/loss function/multilayer perceptron分类
信息技术与安全科学引用本文复制引用
张坤,王贺慈,马金龙,马贵蕾,满梦华,张永强..AR-SNN:脉冲神经网络鲁棒性研究[J].河北科技大学学报,2025,46(5):508-520,13.基金项目
国家自然科学基金(62401623) （62401623）
