华中科技大学学报（自然科学版）2025，Vol.53Issue(10)：15-21,41,8.DOI:10.13245/j.hust.251095
基于统计原理的轻量级混合加权池化方法
Lightweight mixed weighted pooling method based on statistic principle
摘要
Abstract
To address the issue that the impacts of pooling operations in convolutional neural networks(CNNs)on the number of weight parameter updates during network training and the model robustness against adversarial attacks had not been fully investigated,a lightweight mixed weighted pooling method was proposed.Effective features from the median of adjacent features maps were selected and differentiated weight allocation was implemented by this method based on the 3σ principle of normal distribution in statistics.Meanwhile,an independent weighting strategy for max features maps was integrated to enhance key visual information,whereby a theoretically interpretable lightweight pooling mechanism was formed,which significantly reduced the number of weight parameter updates while maintaining feature representation capability.Experimental results show that in the classification tasks of Cifar10 and Cifar100 benchmark datasets on several typical CNNs,substantial compression of the number of weight parameter updates is achieved by the proposed pooling method compared with baseline pooling methods,under the premise that the prediction accuracy is kept close to the optimal level.Additionally,certain improvement in robustness is exhibited by the method in adversarial attack scenarios,outperforming comparative methods under multiple attack modes.关键词
卷积神经网络/池化/权重更新/鲁棒性/对抗样本攻击Key words
convolutional neural networks/pooling/weight update/robustness/adversarial attacks分类
信息技术与安全科学引用本文复制引用
杨美妮,贲可荣,王飞鹏,何涛..基于统计原理的轻量级混合加权池化方法[J].华中科技大学学报（自然科学版）,2025,53(10):15-21,41,8.基金项目
海军装备十四五预研项目(3020904070201) （3020904070201）
海军工程大学自主立项资助项目(2022501040). （2022501040）
