Self-training for unsupervised domain adaptation

Download

thesis.pdf

Date

2022-8-31

Author

Akkaya, İbrahim Batuhan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

204
views

297
downloads

Despite the outstanding performance of deep learning techniques, achieving high per- formance generally demands large amounts of labeled data. Because of the labeling costs, people consider utilizing public datasets or synthetic images with freely gen- erated labels. Unfortunately, deep neural networks are notably sensitive to domain misalignment. The methods to reduce domain misalignment are studied under do- main adaptation (DA). Self-training, which selects a subset of the unlabeled data for pseudo-labeling, has been exploited for DA methods lately. These studies usually ex- ploit a confidence threshold to eliminate inaccurate pseudo-labels. Confidence-based approaches rely on the low-density separation hypothesis, which assumes data is in- dependent and identically distributed. However, the low-density separation hypoth- esis for the target domain for the model trained in the source domain may not hold since the source, and target domains do not share the same distribution. This situation reveals the necessity of a pseudo-labeling metric specific to the target domain. In this thesis, we propose several self-training-based unsupervised DA methods. We evaluate our methods in different modalities such as visible and thermal spectrum, for different tasks such as classification and semantic segmentation, and in different sce- narios such as classical and source-free DA. First, we propose a self-training guided adversarial DA method to promote the generalization capabilities of adversarial DA methods in thermal to RGB modalities. Then, as the main contribution of this thesis, we design a metric learning approach defined in the target domain to enable better guidance for self-training. We use our metric for semantic segmentation tasks in clas- sical and source-free DA scenarios. The experimental results show the superiority of the proposed metric and the effectiveness of the self-training.

Subject Keywords

Domain adaptation, Metric learning, Self-training, Adversarial training

URI

https://hdl.handle.net/11511/99420

Collections

Graduate School of Natural and Applied Sciences, Thesis

Suggestions

OpenMETU
Core

Learning semi-supervised nonlinear embeddings for domain-adaptive pattern recognition Vural, Elif (null; 2019-05-20) We study the problem of learning nonlinear data embeddings in order to obtain representations for efficient and domain-invariant recognition of visual patterns. Given observations of a training set of patterns from different classes in two different domains, we propose a method to learn a nonlinear mapping of the data samples from different domains into a common domain. The nonlinear mapping is learnt such that the class means of different domains are mapped to nearby points in the common domain in order to...
Representation Learning for Contextual Object and Region Detection in Remote Sensing Firat, Orhan; Can, Gulcan; Yarman Vural, Fatoş Tunay (2014-08-28) The performance of object recognition and classification on remote sensing imagery is highly dependent on the quality of extracted features, amount of labelled data and the priors defined for contextual models. In this study, we examine the representation learning opportunities for remote sensing. First we attacked localization of contextual cues for complex object detection using disentangling factors learnt from a small amount of labelled data. The complex object, which consists of several sub-parts is fu...
Domain Adaptation on Graphs via Frequency Analysis Pilancı, Mehmet; Vural, Elif (2019-08-22) Classical machine learning algorithms assume the training and test data to be sampled from the same distribution, while this assumption may be violated in practice. Domain adaptation methods aim to exploit the information available in a source domain in order to improve the performance of classification in a target domain. In this work, we focus on the problem of domain adaptation in graph settings. We consider a source graph with many labeled nodes and aim to estimate the class labels on a target graph wit...
The effects of cognitive load in learning from goal based scenario designed multimedia learning environment for learners having different working memory capacities Kılıç, Eylem; Yıldırım, Zahide; Department of Computer Education and Instructional Technology (2009) The purpose of this study is to investigate the effects of principles aiming to reduce extraneous cognitive load in learning from goal based scenario designed multimedia learning environment for learners having different working memory capacities. In addition, the effects of goal based scenario and the principles of cognitive load theory on students’ perception, motivation and satisfaction has been explored. Two versions of the multimedia were developed for this study. In the first version (+CLT), the princ...
Domain Adaptation with Nonparametric Projections Vural, Elif (2019-08-22) Domain adaptation algorithms focus on a setting where the training and test data are sampled from related but different distributions. Various domain adaptation methods aim to align the source and target domains in a new common domain by learning a transformation or projection. In this work, we learn a nonlinear and nonparametric projection of the source and target domains into a common domain along with a linear classifier in the new domain. Experiments on image data sets show that the proposed nonlinear a...

Citation Formats

İ. B. Akkaya, “Self-training for unsupervised domain adaptation,” Ph.D. - Doctoral Program, Middle East Technical University, 2022.