Pokročilé metody asymetrického heterogenního transfer learningu

Abstract

This dissertation thesis deals with the application of asymmetric heterogeneous transfer learning in scenarios where transfer of data is crucial due to the preservation of the quality of the prediction model. We consider the existence of a source and target domain, where only the source domain is equipped with labels and is thus used to train the model to solve a supervised prediction task which is the same for both domains. In order to solve the same supervised task on an unlabeled target domain one wants to reuse the already trained model. A problem arises when the domains are dierent, i.e. they contain overlapping features with same joint distributions, overlapping features with dierent marginal distributions, or no overlapping features at all. This thesis considers two complex real-world tasks: missing features reconstruction and unsupervised domain adaptation. First, we introduce the reader to the transfer learning eld with comprehensive denitions of the basic concepts. We provide a survey of several methods used within the less known eld of asymmetric heterogeneous transfer learning. Next, we target the missing features reconstruction problem. We focus on a comprehensive survey of the missing data imputation domain and on the possibility of applying these methods to reconstruct entire missing features. We introduce our own imputation model called Wasserstein Generative Adversarial Imputation Network and provide its experimental comparison to state-of-the-art imputation methods. This model is able to solve the considered task based only on a general training strategy even in a scenario when we do not know which features are missing in advance. Finally, we focus on unsupervised domain adaptation. Our aim was to solve the problem of mapping images from one domain to images of another domain without the need of paired or even labeled data. After a survey of the state-of-the-art methods, we propose a novel model called Latent Space Translation Network. This model greatly outperforms other state-of-the-art approaches. This thesis therefore presents two novel methods, which tackle important real-world scenarios.

This dissertation thesis deals with the application of asymmetric heterogeneous transfer learning in scenarios where transfer of data is crucial due to the preservation of the quality of the prediction model. We consider the existence of a source and target domain, where only the source domain is equipped with labels and is thus used to train the model to solve a supervised prediction task which is the same for both domains. In order to solve the same supervised task on an unlabeled target domain one wants to reuse the already trained model. A problem arises when the domains are dierent, i.e. they contain overlapping features with same joint distributions, overlapping features with dierent marginal distributions, or no overlapping features at all. This thesis considers two complex real-world tasks: missing features reconstruction and unsupervised domain adaptation. First, we introduce the reader to the transfer learning eld with comprehensive denitions of the basic concepts. We provide a survey of several methods used within the less known eld of asymmetric heterogeneous transfer learning. Next, we target the missing features reconstruction problem. We focus on a comprehensive survey of the missing data imputation domain and on the possibility of applying these methods to reconstruct entire missing features. We introduce our own imputation model called Wasserstein Generative Adversarial Imputation Network and provide its experimental comparison to state-of-the-art imputation methods. This model is able to solve the considered task based only on a general training strategy even in a scenario when we do not know which features are missing in advance. Finally, we focus on unsupervised domain adaptation. Our aim was to solve the problem of mapping images from one domain to images of another domain without the need of paired or even labeled data. After a survey of the state-of-the-art methods, we propose a novel model called Latent Space Translation Network. This model greatly outperforms other state-of-the-art approaches. This thesis therefore presents two novel methods, which tackle important real-world scenarios.