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 do-mains.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.
transfer learningdomain adaptationmulti-source domain adaptationdeep neural networkdeep learning
万方数据
维普
