Noise Correction Domain Adaptation Learning Based on Classifiers Discrepancy
Unsupervised domain adaption(UDA)aims to transfer knowledge from the related and label-rich source domain to the label-scarce target domain.Usually,domain adaptation methods assume that the source data is correctly labeled.However,the labels and features of source samples will be destroyed due to the actual noise environment.To solve the problem of noisy source domain,this paper proposed noise correction domain adaptation based on classifiers discrepancy(NCDA).First,this method made a more precise classification standard by the difference between multiple classifiers in the network,which can divide noisy source samples into feature noise samples,label noise samples,and clean samples.Second,different correction methods were applied on them.Then,the corrected samples were put back into the training procedure.Finally,this paper used the idea of stochastic classifiers to improve the network.Extensive experiments on Office-31,Office-Home and Bing-Caltech demonstrated the effectiveness and robustness of NCDA,whose accuracy is 0.2%～1.6%higher than the sub-optimal method.
国家科技期刊平台
NSTL
万方数据
维普
