Research On Domain Adaptation Algorithm For Class Correlation Optimization

Machine learning is widely used in real world,but the good performance of machine learning is based on the consistent distribution of training and test sets.In practical application scenarios,it is difficult to satisfy the same distribution of training set and test set,which will bring performance loss.Domain adaptation is proposed to reduce or even eliminate this performance penalty.Research related to domain adaptation has grown rapidly in recent years,and there are different approaches to cope with different problem settings.In this paper,various methods of deep domain adaptation in recent years are summarized,and the domain adaptation algorithm is optimized by class correlation,and finally this optimization is extended to the multi-source domain adaptation range.The main work of this paper is as follows:(1)Reviewed and summarized the deep domain adaptation algorithms in recent years.According to the number of source domains,domain adaptation is divided into single-source domain adaptation and multi-source domain adaptation.According to whether the feature space is shared,it is divided into Homogeneous domain adaptation and Heterogeneous domain adaptation.According to the label set relationship between the source domain and the target domain,is divided into closed set case,open set case,partial set case,universal case and zero-sample case.Among them,closed-set domain adaptation is divided into difference-based methods,confrontation-based methods and self-training-based methods according to the core idea.(2)Proposed adversarial domain adaptive algorithm via class correlation optimization.The accuracy on the public datasets Office-31,Office-Home and Vis DA2017 generally exceeds the listed methods,and the convergence speed is also excellent.This shows that the method in this paper can reach the convergence condition under the condition of limited computational performance.Minimizing the entropy of the class correlation matrix can be transformed into maximizing the kernel norm of the class correlation matrix.By maximizing the kernel norm,the class correlation matrix can be given more diversity,making the entire network more robust.(3)Finally,for multi-source domain adaptation,this paper measures the distribution difference between the two domains through the feature norm difference between the source domain and the target domain.Narrows the difference between the source domain feature norm and the target domain feature norm in the iterative process,and the distribution difference between the two domains is reduced.In order to avoid the gradient explosion caused by the norm greater than 1 in the back-propagation process,the iterative update is adopted for the feature norm and does not participate in the gradient calculation.By introducing the class correlation loss,the average performance of the multi-source domain adaptation algorithm optimized by the class correlation loss is improved by 3.2% on the Office-Home dataset and 2.9% on the Domain Net dataset.