Study Of Robust Domain Adaptation Algorithm With Label Noise

In recent years,deep learning has relied on efficient computing devices,powerful algorithms and vast amounts of data to achieve impressive results in areas such as computer vision,data mining,speech recognition and healthcare.The success of deep learning relies on massive amounts of training data,yet obtaining large-scale annotated data is not easy,costly and time-consuming.At the same time,the distribution of data varies across scenarios,and models trained on a particular scenario often perform poorly in other scenarios.Therefore,it is an important and challenging problem to make full use of existing data to improve the generalisation ability of models and to compensate for the lack of samples.To address these issues,transfer learning has been developed.Domain adaptation is a subfield of transfer learning,which transfers knowledge learned in the source domain using labelled samples to the target domain,thereby improving the task performance of the target domain.There have been many achievements in domain adaptation research,however,there are still two problems : 1)the current mainstream approaches are based on the ideal assumption of accurate labeling of source domain data,which is difficult to meet in the real world;2)the current mainstream approaches tend to use the same fixed threshold for each category when using target domain pseudo-labels,without considering the impact of confidence on accuracy and category balance,which may lead to the loss of some category samples during the screening process and thus affect the migration effect.In this work,the Weakly Supervised Variational Adversarial Domain Adaptation(WSVADA)algorithm is proposed to address these two problems.The algorithm is divided into two main phases: pre-training and domain adaptation.In the pre-training phase,we use a two-view classifier to output confidence ambiguities and filter out samples with consistent judgments to participate in the training,so as to achieve noise reduction on the source domain data.In the domain adaptation phase,we propose a dynamic threshold filtering algorithm(DTFA)that takes into account the category balance to assign pseudo-labels to the target domain samples,extract the target domain features using a variational self-encoder,and approximate the target domain category-level features to the source domain by maximizing the variational lower bound;then use a generative adversarial network The global features of the source domain are then approximated to the target domain using a generative adversarial network.This algorithm effectively mitigates the adverse effects of source domain label noise and pseudo-label noise on the domain adaptation task,and further facilitates the alignment of category-level features between the two domains.In this paper,six sets of migration experiments were conducted on the Office-31 dataset with two different category-noise treatments,and the average accuracy of the experimental results reached the leading level in the field,demonstrating the effectiveness of the algorithm.