Unsupervised Domain Adaptation with Robust Deep Logistic Regression

Abstract

The goal of unsupervised domain adaptation (UDA) is to eliminate the cross-domain discrepancy in probability distributions without the availability of labeled target samples during training. Even recent studies have revealed the benefit of deep convolutional features trained on a large set (e.g., ImageNet) in alleviating domain discrepancy. The transferability of features decreases as (i) the difference between the source and target domains increases, or (ii) the layers are toward the top layer. Therefore, even with deep features, domain adaptation remains necessary. In this paper, we treat UDA as a special case of semi-supervised learning, where the source samples are labeled while the target samples are unlabeled. Conventional semi-supervised learning methods, however, usually attain poor performance for UDA. Due to domain discrepancy, label noise generally is inevitable when using the classifiers trained on source classifier to predict target samples. Thus we deploy a robust deep logistic regression loss on the target samples, resulting in our RDLR model. In such a way, pseudo-labels are gradually assigned to unlabeled target samples according to their maximum classification scores during training. Extensive experiments show that our method yields the state-of-the-art results, demonstrating the effectiveness of robust logistic regression classifiers in UDA.