Abstract

In today's world, the need to analyze and classify data (images, videos and text) is felt more and more due to the rapid growth of the information volume through the applications and social media. Machine learning and transfer learning are the most widely used branches of artificial intelligence to identify and classify data. For training a machine learning model to classify data, a sufficient number of labeled data is required, but unfortunately finding labeled data is often time consuming and costly and requires human skills. Therefore, to classify the unlabeled data, a model is trained on labeled and accessible data, and the gained knowledge is used to label the unlabeled data. However, the machine learning methods are useful for problems where the source and target data have the same feature space with same distribution. Thus, if there is a discrepancy in the feature space and / or distribution of data, for example in the field of speech processing, source and target data are different in terms of background noise, tone of voice and gender of the speaker, the trained model on source data cannot be used on target data, easily. In this case, there are two solutions: 1) reconstruct the model (classifier) for each new condition (data), 2) adapt the source and target representations to use the knowledge from the source domain to classify the target domain, which is known as domain adaptation or transfer learning. In this thesis, three unsupervised transfer learning methods and two semi-supervised transfer learning methods are suggested. In first, a new two-step approach is proposed in unsupervised transfer learning manner that called &ldquooptimal Classifier learning with weighted Geometric Mean and Dynamic balance distribution alignment (CGMD)&rdquo. In the first step, the CGMD method uses the weighted geometric mean of the second-order statistics of the source and target domains to adapt the source and target domains. In second step, the CGMD learns an optimal classifier by the structural risk minimization and dynamic balance distribution alignment in the Reimannian manifold space. &ldquoAdaptive Classifier Learning via Two-step MMD and Exploiting domain structures (ACLTE)&rdquo is the second proposed method. ACLTE introduces modified visual domain adaptation, which utilizes balanced maximum mean discrepancy (MMD) to perform better domain adaptation. Also, for learning robust classifier against domain shift, set of graph manifold regularizer and modified joint probability maximum mean discrepancy are simultaneously exploited to capture the domain structures and adapt the distributions of projected samples during the model learning process. As the third method, we introduce &ldquoJoint Distinct Subspace learning and unsupervised transfer Classification for visual domain adaptation (JDSC)&rdquo, which is an iterative two-step framework. JDSC is based on hybrid of feature-based and classifier-based approaches that uses the feature-based techniques to tackle the challenge of domain shift and classifier-based techniques to learn the reliable classifier. In addition, for subspace alignment, weighted joint geometrical and statistical alignment is proposed to learn two coupled projections for mapping the source and target data into respective subspaces by accounting the importance of the marginal and conditional distributions, differently. The fourth method, entitled "Latent Sparse subspace learning and visual domain classification via Balanced distribution alignment and Hilbert-Schmidt metric (LSBH)" is a semi-supervised method for transfer learning. The LSBH method uses simultaneous latent space learning and sparse reconstruction to adapt domains. To overcome the overfitting problem in classifier learning, the sparse constraint is utilized for the reconstruction matrix, which selects a small amount of source data for knowledge transfer. Simultaneous latent subspace learning and sparse reconstruction prevents the creation of local common subspace for source and target domains to achieve global optimal common subspace. Also, for obtaining an optimal classifier, label prediction loss is reduced. In addition, two criterias (i.e., maximum mean discrepancy and Hilbert-Schmidt independence criteria) are used to reduce the marginal and conditional distribution disparities of domains and increase the dependence between samples and labels for optimal classifier finding. In order to maintain the geometric structure of data in classification step, the neighborhood graph of the samples is used. We proposed our fifth method, called &ldquoKErnelized Domain Adaptation and balanced distributions alignment for image classification (KEDA)&rdquo to adapt the source and target domains in a semi-supervised manner. KEDA preserves the topology of domains via creating a Laplacian matrix and similarity and dissimilarity views. Moreover, KEDA adapts the regularized conditional and marginal distributions across domains by finding the optimal classification function. Also, at the stage of finding the optimal objective function, it minimizes the structural risk of the function on source domain. Ultimately, the sum of these solutions leads to a good classification function for labeling the unlabeled images.

Key Words: Image classification, Domain adaptation, Manifold matching, Balanced distribution adaptation, Transfer learning, Distance metric