一种基于伪标签的深度迁移学习算法  

A Deep Migration Learning Slgorithm Based on Pseudo-labeling

作　　者：葛一正 李云栋[1] 

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《工业控制计算机》2023年第11期131-133,共3页Industrial Control Computer

基　　金：国家自然科学基金资助项目(62071006)。

摘　　要：深度学习的成功依赖于海量的训练数据,然而获取大规模有标注的数据并不容易,成本昂贵且耗时;同时由于数据在不同场景下的分布有所不同,利用某一特定场景的数据集所训练出的模型往往在其他场景表现不佳。迁移学习作为一种将知识从一个领域转移到另一个领域的方法,可以解决上述问题。深度迁移学习则是在深度学习框架下实现迁移学习的方法。提出一种基于伪标签的深度迁移学习算法,该算法以ResNet-50为骨干,通过一种兼顾置信度和类别平衡的样本筛选机制为目标域样本提供伪标签,然后进行自训练,最终实现对目标域样本准确分类,在Office-31数据集上的三组迁移学习任务中,平均准确率较传统算法提升5.0%。该算法没有引入任何额外网络参数,且注重源域数据隐私,可移植性强,具有一定的实用价值。The success of deep learning relies on massive amounts of training data.However,obtaining large-scale labeled data is not easy,it is expensive and time-consuming.At the same time,because the distribution of data varies across scenarios,models trained using datasets from a particular scenario often perform poorly in other scenarios.Transfer learning,as a method to transfer knowledge from one domain to another,can solve the above problems.Deep transfer learning,on the other hand,is a method to implement transfer learning in a deep learning framework.This paper proposes a deep transfer learning algorithm based on pseudo-labeling,which uses ResNet-50 as the backbone,provides pseudo-labeling for target domain samples through a sample selection mechanism that takes into account confidence and category balance,and then conducts self-training to finally achieve accurate classification of target domain samples.In three sets of transfer learning tasks on the Office-31 dataset,average accuracy improvement of 5.0%over traditional algorithms.This algorithm does not introduce any additional network parameters and focuses on the privacy of source domain data.

关 键 词：深度学习 迁移学习 伪标签 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]