Unsupervised Transfer Softmax Regression

Abstract

Cross-domain image classification is a challenge in numerous practical applications due to the variance between the training and testing datasets. To solve the problem, we propose a new classification method named unsupervised transfer softmax regression in this paper. It firstly introduce joint distribution adaptation to the objective function of the softmax regression to construct a new classifier for knowledge transfer. Then the new objective function is solved by gradient descent method to realize the unified optimization of classification and feature extraction. Finally, we evaluate the effectiveness of the proposed method by the classification experiments on image data sets and text data sets, and the result demonstrate the good performance of our approach.