A two-branch symmetric domain adaptation neural network based on Ulam stability theory

Abstract

Unsupervised domain adaptation is an important branch of transfer learning, which mainly solves the generalization problem of models from fully labeled datasets to unlabeled datasets. Unsupervised domain adaptation mainly aims to make full use of domain features, suppress inter domain distribution shift, and improve the domain generalization ability of the model. In this paper, we propose a two-branch symmetric domain adaptation neural network. In each independent domain adaptation branch, we use the method of adversarial learning to realize the embedding and alignment of domain features, and use a new generalized autoencoder model based on Ulam stability theory to enhance the feature extraction ability of the network. A fusion network is adapted to fuse the features extracted from each branch of the feature subspace to further improve the domain adaptability and generalization ability of the model. Experimental results show the effectiveness of the proposed method.