Domain adaptation network based on autoencoder

Abstract

The domain adaptation uses labeled source domain data to train a classifier to be used in the target domain with no or small amount of labeled data. Usually there exists discrepancy in terms of marginal and conditional distributions for both source and target domains, which is of critical importance to minimize the distribution discrepancy between domains. As a classical model in deep learning, the autoencoder is capable of realizing distribution matching and enhancing classification accuracy by extracting more abstract and effective features from data. A Domain adaptation network based on autoencoder (DANA) is proposed. The DANA structure consists of a couple of encoding layers: a feature extraction layer and a classification layer. For the feature extraction layer, the marginal distributions of source and target domains are matched by using the nonparametric maximum mean discrepancy measurement. For the classification layer, the softmax regression model is applied to encode the label information of source domains meanwhile to match the conditional distribution. Experimental results on ImageNet, Corel and Leaves datasets have shown the enhanced classification accuracy by our proposed algorithm compared with the classical methods.