Multi-source Deep Transfer Learning Algorithm Based on Feature Alignment

With the deepening of transfer learning research, researchers are no longer satisfied with the classification of knowledge in a single field but hope that the classification of knowledge in multiple fields can be realized, so as to simulate the behavior of human “analogy” and enable the machine to draw inferences”. However, the feature realization of multiple source domains often differs greatly, which brings a challenge to the traditional transfer learning scheme. In this paper, a multi-source deep transfer learning algorithm MDTLFA based on feature alignment is proposed to solve the problem that the data from multiple source domains often has different feature realizations. MDTLFA first reduces the difference in the marginal probability distribution between fields at the sample level by means of the maximum mean deviation MMD. Then, the feature alignment strategy is used at the feature level to further reduce the difference in the marginal probability distribution between the fields and maintain the unique data manifold structure while sharing similar features. On this basis, the conditional probability adaptation CPDA was constructed to reduce the difference in conditional probability distribution between domains and enhance the portability of source domain features. The CPTCNN model was constructed based on a convolutional neural network using CPDA. Finally, the CPTCNN model is trained in the subspace to obtain a classifier set, and the designed strategy is used to select the classifier with a small classification error in the target domain to form MDTLFA. Multiple source domains, marginal probability adaptation at the sample level and feature level, and the CPTCNN model constructed based on the minimization of conditional probability differences effectively improve the performance of data features in multiple domains, thus improving the classification effect. The experimental results on several real data sets show that the MDTLFA algorithm is effective and has some advantages compared with the advanced benchmark algorithm.