选择置信伪标签的迁移学习
The Transfer Learning via Selecting Confident Pseudo-Labels
滕少华 1周德根 1滕璐瑶 2张巍1
作者信息
- 1. 广东工业大学计算机学院,广东 广州 510006
- 2. 广州番禺职业技术学院信息工程学院,广东 广州 511483
- 折叠
摘要
域适应旨在将标签丰富的源域知识迁移到无标签的目标域.选择性伪标签和标签传播都是域适应的常用方法.然而传统的选择性伪标签以最大类的预测概率标记样本,忽视了其他概率;而且传统的标签传播同等对待不同置信度的标签,这可能导致错误标签.因此,该文提出了一种选择置信伪标签(TL-SCP)的迁移学习.首先,在评估伪标签的置信度时兼顾了最大类的预测概率和其他类预测概率;其次,在标签传播过程中尽量保留高置信度标签,并据此指导低置信度标签的更新,借此减少错误标签传播;最后,在 4 个基准数据集上进行的大量实验验证了提出的模型(TL-SCP)优于现有的模型.
Abstract
Domain adaptation aims to transfer knowledge from well-labeled source domain to unlabeled target domain.Selective pseudo-labels and label propagation are common methods of domain adaptation.The existing methods have the following drawbacks.On the one hand,traditional selective pseudo-label classifies samples with the largest predicted probability of class and ignoring other probabilities.On the other hand,traditional label propa-gation equally treats labels with different confidence,which may lead to mislabeling.Therefore,the transfer learning via selecting confident pseudo-labels(TL-SCP)is proposed.Firstly,when evaluating confidence of pseudo-labels,the maximum prediction probability of the class and the prominence of other prediction probabilities are computed.Secondly,label propagation keeps high confidence labels and let them guide the update of low-confidence labels,so as to reduce the propagation of false label.Finally,a large number of experiments on four benchmark datasets vali-date the proposed model(TL-SCP)over existing advanced models.
关键词
置信伪标签/域适应/伪标签/迁移学习/标签传播Key words
confidence pseudo-label/domain adaptation/pseudo-label/transfer learning/label propagation引用本文复制引用
出版年
2024
NETL
NSTL
万方数据