多源域适应方法综述

A Review of Multi-source Domain Adaptation

域适应是解决源域样本和目标域样本不满足独立同分布问题的迁移学习范式,是当下研究的重点方法。然而实际情况下获取源域样本的渠道和方法并不唯一,这会导致源域中存在多种不同分布的样本。多源域适应方法是解决源域样本分布多样性问题的有效途径,其主要研究各源域分布间的关系和与目标域分布对齐的策略,进一步减轻各域之间的域偏移,具有实用意义和挑战价值。随着深度学习技术的不断进步,多源域适应方法主要使用深度神经网络提取各域的域不变特征作为分布对齐的依据,结合使用度量准则衡量分布差异或者利用对抗思想对齐域间分布。经过理论证明和实验验证,多源域适应方法训练的模型比单源域方法训练的模型具有更好的泛化性能,更符合现实需求。通过介绍多源域适应的研究现状和相关概念,对现有算法进行总结和综述,按照迁移方式不同对多源域适应方法进行分类,进一步分析多源域适应方法性能的实验结果,阐述其存在的不足和缺点,并对多源域适应领域的发展和趋势进行预测。

Domain adaptation is a current research paradigm in transfer learning that solves the problem of non-i.i.d.(independent and identically distributed)source and target domain samples.However,in practice,there are multiple channels and methods for obtaining the source domain samples,which can lead to various distributions within the source domain.Multi-source domain adaptation is an effective approach to address the problem of diversity of source domain sample distributions.It mainly studies the relationships among different source domain distributions and the alignment strategies with the target domain distribution,further reducing the domain shift between domains.This has practical significance and challenging values.With the continuous advancement of deep learning technology,multi-source domain adaptation methods mainly use deep neural networks to extract domain-invariant features from each domain as the basis for distribution alignment.These methods combine the use of metric criteria to measure distribution differences or use adversarial ideas to align domain distributions.Through theoretical proof and experimental verification,models trained by multi-source domain adaptation methods have better generalization performance than that of single-source domain methods and are more in line with real-world needs.We summarize and review existing algorithms for multi-source domain adaptation by introducing the research status and related concepts of multi-source domain adaptation.The methods are classified according to different transfer modes,and the experimental results of their performance are further analyzed.The shortcomings and deficiencies of these methods are also discussed,and predictions are made on the development and trends in the field of multi-source domain adaptation.