一种基于伪标签的半监督少样本学习模型

A Few-Shot Learning Model Based on Semi-Supervised with Pseudo Label

如何将带有大量标记数据的源域知识模型迁移至带有少量标记数据的目标域是少样本学习研究领域的热点问题.针对现有的少样本学习算法在源域数据与目标域数据的特征分布差异较大时存在的泛化能力较弱的问题,提出一种基于伪标签的半监督少样本学习模型FSLSS(Few-Shot Learning based on Semi-Supervised).首先,利用pytorch深度学习框架建立一个关系型深度学习网络,并使用源域数据对网络进行预训练;然后,使用此网络对目标域数据进行分类预测,将分类概率最大的类标签作为数据的伪标签;最后,利用目标域的伪标签数据和源域的真实标签数据对网络进行混合训练,并重复伪标签标记与混合训练过程.实验结果表明,相对于现有主流少样本学习算法,FSLSS模型有更好的泛化能力及知识迁移效果.

How to migrate a source domain knowledge model with a large amount of tagged data to a target domain with a small amount of tagged data is a hot issue in few-shot learning.For the problems that the existing few-shot learning algorithm have weak generalization ability when the difference between the feature distribution of the source domain data and the target domain data is large,a few-shot learning model based on semi-supervised FSLSS(Few-Shot Learning based on Semi-Supervised) is proposed.Firstly,a relational deep learning network is established by using the pytorch framework,and the network is pre-trained by the source domain data.Then,the network is used to predict the target domain data,and the label with the highest classification probability is used as the data’s pseudo label.Finally,the network is hybrid trained using the pseudo label data of the target domain and the real label data of the source domain,then repeating the pseudo-labeled and hybrid trained process.The experimental results show that the FSLSS model has better generalization ability and knowledge transfer effect than the existing few-shot learning algorithms.