Learning to adapt deep neural networks towards a novel image domain with a few target labels

Abstract

In recent years, image classifiers based on deep neural networks have received a great attention due to the outstanding performance. However, the strong dependency on training resources makes the data-driven image classifiers vulnerable to variations of image domain characteristics. One simple yet naive solution for resolving this issue is to mining a large number of labeled images for every image domain of interest, but tremendous expenses are compelled for this approach. To address this problem via an algorithmic approach, in this thesis, I present a comprehensive study on domain adaptation (DA) for deep learning-based image classifiers. DA is a transfer learning methodology that aims to enhance the performance of deep neural networks in a label-scarce domain (target domain) by leveraging knowledge in a label-sufficient domain (source domain). Among various DA categories, I focus on semi-supervised domain adaptation (SSDA) and few-shot supervised domain adaptation (SDA). Both tasks commonly assume that a large number of labeled source images are given for training, and the two tasks are different in the composition of accessible data in the target domain. First, SSDA assumes that a large number of target images are accessible for training, but only a few examples among them are labeled. For SSDA, I propose a selective pseudo labeling scheme for assigning pseudo labels of high reliability to unlabeled target images. In addition, based on the observation that pseudo labels are inevitably noisy, a label noise-robust learning scheme is applied, which alternately updates the network and the set of pseudo labels. Second, SDA considers the case that only a small number of labeled images are available for training without any additional unlabeled image. For SDA, deep features are extracted and projected towards a hyper-spherical space via l2-normalization. Afterwards, an additive angular margin loss is embedded so that the features of both domains are compactly grouped on the basis of the shared class prototypes. To further reduce the domain discrepancy, a pairwise spherical feature alignment loss is incorporated. Experimental validations were conducted on various DA datasets and the results demonstrate that both of the proposed SSDA and SDA methods outperform other previous state-of-the-art methods.