Unsupervised Deep Domain Adaptation Based on Weighted Adversarial Network

Recent studies indicate that adversarial learning can reduce distribution discrepancy between domains effectively, but when the samples belonged to different classes have similar characteristics in the domains, they may be incorrectly aligned to similar classes after domain adaption, which gives rise to negative transfer. To prevent such misalignment, we propose a weighted adversarial network based unsupervised domain adaptation method. Its contributions are mainly reflected in the following two aspects: 1) according to the similarity of features between classes, the different weights are given to the corresponding domain discriminators, which means that we will focus on the alignment of the classes with similar characteristics; 2) all domain discriminators will be given a certain weight again based on the entropy of true or pseudo label vectors, that is, the clearer the sample classification result, the greater its credibilty during domain discriminators learning. Experimental results on several cross-domain benchmark data sets show that our newly proposed approach outperforms state of the art methods.