Transfer Weight Based Conditional Adversarial Domain Adaptation

Domain adaptation is a machine learning paradigm that solves the problem of independent and identical distribution of training data and test data.In recent years,with the development of adversarial learning,the domain adaptation based on adversarial learning has received great attention.Existing domain adaptation methods based on adversarial learning often to align the source and target domains perfectly in the feature space.However,in the actual application scenario,there are some samples that are hardto-transfer in the source domain.Transfer of these samples will disturb the distribution of the target domain samples.In this paper,we propose a transfer weight based conditional adversarial domain adaptation(TW-CDAN)for the failure of the conditional adversarial domain adaptation(CDAN)to fully utilize the sample transferability,which still struggle with some hard-to-transfer source samples disturbed the distribution of the target domain samples.The main work of this paper are following:In order to better to excavate the transferability of the sample,the domain discriminant result is used as the main factor of transferability.First,the transfer weight of the sample is calculated by the entropy of the domain discrimination model's result.Secondly,the transfer weight is applied to the classification loss of the source domain sample to control the samples of different transfer weights,and different classification loss weights are used to update.The network aims to eliminate the impact on hard-totransfer samples on the model.In order to better to solve the problem of pattern collapse in the adversarial learning.The minimum entropy loss based on Transfer weight is further introduced,and the separation boundaries of different classes of the source domain sample sets are increased to reduce the pattern collapse problem faced by the adversarial model.Compare with existing domain adaptation methods of 24 transfer tasks under two data sets.Experiments have shown that TW-CDAN has better performance.