基于领域困惑损失函数的迁移学习算法  

Transfer Learning Algorithm Based on Domain Confusion Loss Function

作　　者：王孝顺 于雅静 WANG Xiaoshun;YU Yajing(Huzhou College,Huzhou 313000;Hunan Technology and Business University,Changsha 410205)

机构地区：[1]湖州学院,湖州313000 [2]湖南工商大学,长沙410205

出　　处：《计算机与数字工程》2024年第9期2771-2777,共7页Computer & Digital Engineering

摘　　要：现实中针对目标任务往往没有标签数据,如何在辅助源领域中学习到知识以应用目标任务成为了一个关键性问题。为此提出一种新方法,即领域困惑迁移网络(Domain Confusion Transfer Network,DCTN)。该方法首先在残差网络的基础上设计出领域分布部分,然后在领域分布部分的输出端使用领域困惑损失函数以使得特征提取网络学习领域不变特征。通过实验证实,该方法可以有效地学习领域不变特征且能够获得良好的效果。In reality,there is often no labeled data for the target task,and how to learn knowledge in the auxiliary field to apply the target task has become a key issue.A new method,namely domain confusion transfer network(DCTN),is proposed.This method first designs discriminator network on the basis of residual network,and then uses domain confusion loss function on the output of domain distribution part to make the feature extraction network learning the domain invariant feature.It is proved by experiment that this method can effectively learn domain invariant feature and achieve good results.

关 键 词：标签数据 目标任务 残差网络 领域分布 领域困惑损失函数 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]