| Domain adaptive(DA)learning,which aims to utilize label-rich source domains to help labelscarce target domain learning,has received much attention in recent years.Considering feature distribution difference between the source domain and the target domain,existing DA methods often use a distance-based or adversarial-based approach to align feature distribution and learn domaininvariant features,thereby transferring the class discrimination knowledge in the source domain to the target domain.DA methods usually assume that the source domain data is properly labeled,however,it is often difficult to collect a large number of source instances with clean labels in real tasks.Label-noisy DA learning with noisy source labels may degrade target performance,incurring negative transfer.Therefore,this paper studies the domain adaptation scenario where the source domain contains label noises.The main research contents include the following two parts:First,a noise-robust conditional domain adaptation method Ro CDAC based on sample resampling and noise label correction is proposed.Ro CDAC first re-weight samples through metanetwork,so that the low-loss samples get higher weights and contributions in classification.Then,the noisy samples are separated from the clean samples based on the threshold method,and the labels of noisy samples are corrected by the prototype network and re-invested in adaptation.Experiments show that the proposed method achieves a significant performance improvement compared to the state-of-art DA method on the manually polluted benchmark datasets.Second,a noise-robust adversarial domain adaptation method Ro ADAS with prototype-based label noise detection and robust regularization is proposed.The method uses Gaussian Mixture Model(GMM)to perform label noise detection on noisy source samples in a class-independent manner based on prototypes.A regularization loss is then introduced to increase the distribution distance between the model’s predicted probability output and noisy labels.Finally,comparative experiments are carried out on the manually polluted DA benchmark dataset to verify the effectiveness of the method.All the strategies proposed in this paper can be integrated into existing domain adaptation methods to improve their performance in label noise DA scenarios with considerable flexibility. |
