基于联邦加权域自适应的多站点自闭症诊断
Federated weighted domain adaptation for multi-site autism diagnosis
郝小可 1甄时伟 1周超飞 1马明明 1刘时宇 1曹家辉1
作者信息
- 1. 河北工业大学 人工智能与数据科学学院,天津 300401
- 折叠
摘要
为了提高联邦学习(Federated Learning,FL)方法在多站点自闭症诊断中对数据异质性的关注和解决部分站点缺乏标注数据导致的局部阻塞问题,提出了联邦加权域自适应(Federated Weighted Domain Ad-aptation,FedWDA)方法.首先,不同于聚合一个共享的全局模型的方法,FedWDA在模型聚合的过程中保留了每个本地模型的批量归一化(Batch Normalization,BN)层,使用推土机距离(Earth Mover's Dis-tance,EMD)计算了BN层的统计信息间的相似性,用于指导站点上模型的加权聚合,为每个站点都提供了更加个性化的本地模型,缓解了数据异质性导致的分类准确度不足的问题;其次,对于缺乏数据标注的站点,FedWDA基于样本伪标签的一致性对本地模型进行了无监督聚类调整,充分利用了无标注站点的数据.在公开的数据集ABIDE上的实验结果表明,所提方法相比其他传统方法在多站点自闭症诊断中具有更高的准确率.
Abstract
To address the lack of attention to data heterogeneity and solve the local blocking problem caused by the lack of labeled data at some sites in Federated Learning(FL)multi-site autism diagnosis methods.Federated Weighted Domain Adaptation(FedWDA)has been proposed.Firstly,instead of aggregating a global model,FedWDA retains the Batch Normalization(BN)layer for each local model and calculates the similarity between the BN layer statistics with Earth Mover's Distance(EMD)to guide the models'weighted aggregation.Each site is provided with a more personal-ized local model,alleviating the problem of poor accuracy caused by data heterogeneity.Secondly,unsupervised cluster learning is used to fine-tune the local models of sites lacking labeled data,based on the pseudo-label's consistency.The experimental results on ABIDE show that the proposed method can perform multi-site autism diagnosis more accurately than other traditional methods.
关键词
联邦学习/多站点/领域自适应/自闭症诊断/聚类学习Key words
federated learning/multi-site/domain adaptation/autism diagnosis/clustering learning引用本文复制引用
基金项目
国家自然科学基金资助项目(62276088)
河北省自然科学基金资助项目(F2023202072)
出版年
2024
NETL
NSTL
万方数据