Research On Noisy Domain Adaptation Based On Correction Mechanism

Unsupervised Domain Adaption(UDA)aims to transfer knowledge from the related and labelrich source domain to the label-scarce target domain.Usually,current mainstream domain adaptation methods assume that source data is correctly labeled,and the label set of source domain and target domain is consistent.However,in real tasks,noisy environment can damage labels and features of source domain samples,making it difficult to collect a large amount of labeled clean source data.Extracting common knowledge between domains from noisy source domain is a huge challenge for in noisy domain adaptation.At the same time,there may be inconsistencies in label sets between domains.Each domain often contains non-shared private classes,which seriously affects the learning performance of the target domain.For this reason,this paper focuses on the research of standard domain adaptation with consistent label sets and universal domain adaptation with mismatched label sets with noisy source domain.The main content of this paper includes the following two parts:First,a new method named Dual-Correction Adaptation Network for Noisy Knowledge Transfer(Dual CAN)is proposed.Dual CAN adopts a dual-directional network to help both source and target domain transfer knowledge from each other and correct noise labels.In addition to inherent noise in source domain,Dual CAN regards low confidence label in target domain as noise and adds a noise recognition and correction module to the dual network to correct label noise and feature noise in both domains,simultaneously improving the learning performance of source and target domains.Second,the other method named Noise Correction Universal Domain Adaptation based on Classifiers Discrepancy(NCUDA)is proposed.According to the discrepancy between two classifiers,this method divides source data into three subsets: clean sample set,hard sample set,and noisy sample set.For hard and noisy samples,NCUDA uses weight adaptation and correction to make them close to the corresponding category center,and then puts them back into the training procedure with clean source samples.After that,the discrepancy of classifiers with source knowledge is used to filter out target private samples.The whole network transfers common knowledge from source to target in an adversarial way.Finally,NCUDA uses the idea of stochastic classifiers to improve the network.Extensive experiments on Office-31,Office-Home and Bing-Caltech demonstrats the effectiveness and robustness of Dual UAN and NCUDA under different noise rates.