Correlation alignment for domain adaptatio

Unlike human learning, machine learning often fails to handle changes between training (source) and test (target) input distributions. Such domain shifts, common in practical scenarios, severely damage the performance of conventional machine learning methods. Supervised domain adaptation methods have been proposed for the case when the target data have labels, including some that perform very well despite being frustratingly easy to implement. However, in practice, the target domain is often unlabeled, requiring unsupervised adaptation.;This thesis proposes a simple, effective, and efficient method for unsupervised domain adaptation called CORrelation ALignment (CORAL). CORAL minimizes domain shift by aligning the second-order statistics of source and target distributions, without requiring any target labels. It is more general than the subspace manifold methods because it aligns the distributions of the source and target domain rather than the bases. It is also much simpler than other distribution matching methods. CORAL performs remarkably well in extensive evaluations on standard benchmark datasets.;We also develop novel domain adaptation algorithms by applying the idea of CORAL to three different scenarios. For linear classifiers, we equivalently apply CORAL to the classifier weights, leading to added efficiency when the number of classifiers is small but the number and dimensionality of target examples are very high. The resulting CORAL Linear Discriminant Analysis (CORAL-LDA) outperforms LDA by a large margin on standard domain adaptation benchmarks. The second scenario is applying CORAL to subspace manifold based methods.;We incorporate the idea of CORAL into two recently published methods and the resulting CORAL-SS method outperforms its counterparts consistently. Last but not least, we extend CORAL to learn a nonlinear transformation that aligns correlations of layer activations in deep neural networks. The resulting Deep CORAL approach works seamlessly with deep networks and achieves state-of-the-art performance on standard benchmark datasets.;The last part of the thesis explores two new scenarios for visual domain adaptation: from virtual to reality and from ground to sky. In virtual to reality, the source domain contains synthetic images generated from 3D CAD models while the target data are real images. For ground to sky, the target domain includes aerial images while the source data are the common consumer photos. These new scenarios provide much larger shifts and more practical applications to facilitate future research in domain adaptation.