Research And Application Of Multi-source Domain Adaptation Method

With the rapid development of market economy and science and technology,driven by the powerful engine of artificial intelligence,a large number of machine learning application scenarios have emerged.The good performance of supervised learning requires a large amount of labeled data.This task is expensive and tedious.Therefore,transfer learning is receiving more and more attention in the intelligent era.Among the related technologies of transfer learning,multi-source domain adaptation is one of the important key technologies.Multi-source domain adaptation research how to use the knowledge of multiple source domains to learn the target domain,strengthen domain adaptation capabilities,and alleviate negative transfer.In recent years,thanks to the rapid development of artificial intelligence and computer vision,certain results have also been achieved in multi-source domain adaptation image classification technology.However,there are still many difficulties that have not been overcome,such as feature selection and relationship matching techniques between various source domains,feature adaptation and relationship matching algorithms between multiple source domains and target domains,and multi-source domain adaptation in the computer Vision and other applications.Therefore,in-depth research on adaptation methods in multi-source domains,research and development of new multi-source domain adaptation algorithms,and exploration of new multi-source domain adaptation application technologies are of great significance.Based on the above research background,this dissertation focuses on multi-source domain adaptation technology.By analyzing the status quo of adaptation research in multi-source domains at home and abroad,and aiming at the difficulties in the research of adaptation methods in multi-source domains,a new multi-source domain adaptation technology method and application scheme are proposed.The following results have been achieved:(1)This thesis studies the relevant algorithms and theoretical methods of multi-source domain adaptation,and systematically organizes and classifies the existing algorithms.Review the current research work of multi-source domain adaptation for each category,and give a standardized description and schematic diagram.(2)This thesis proposes a relationship-based adaptation method for multi-source domains.Use cycles and duality to connect various domains.The data of the source domain and the target domain are related through a certain similarity.That is,the data of one domain is transferred to another domain through the inner product form of the embedded space,and then it is transferred back to form a cycle.This method achieves a better domain adaptation effect when using a small amount of resources.(3)This thesis proposes a confrontation-based sub-domain feature alignment domain adaptation method.Using the method of generating a confrontation network,each domain shares a common feature extractor,and then the domain aligns the source domains first according to some conditions(such as category labels),and divides similar samples into the same subdomain to reduce the difference between multiple sources Intra-class distance,increase the inter-class distance.This method has achieved high accuracy results in experiments.(4)This thesis proposes a multi-source domain adaptation model based on the risk investigation and classification of judicial cases.Multi-source domain adaptation judicial application in various courts under City A.Through data migration and model fine-tuning,on the basis of existing data,adaptation applications in multi-source domains have achieved better results.