Year: 2023
Type: dissertation
Abstract: Transfer learning is a prominent paradigm in machine learning, addressing the challenges of model generalization across different domains. This dissertation delves into two fundamental aspects: domain generalization and domain adaptation. In the context of domain generalization, the objective is to build models capable of effectively generalizing to unseen target domains by leveraging information from multiple source domains during training. Conversely, domain adaptation focuses on adapting a model trained on a specific source domain to a distinct target domain, assuming the availability of data from the target domain during training. Furthermore, this work explores close-set and open-set settings in transfer learning. The close-set setting assumes a fixed set of known target classes, while the open-set setting handles the challenge of novel, unseen classes in the target domain during inference. Understanding these settings' nuances is crucial for developing versatile transfer learning strategies that cater to diverse real-world scenarios and data distributions. The primary goals of transfer learning in this study include enhancing the model's performance on the target domain and reducing training costs by leveraging knowledge from a related source domain. However, achieving these goals comes with challenges, including addressing domain shift, preventing negative transfer, and handling unknown classes in open-set settings. By investigating these challenges and developing innovative transfer-learning techniques, this dissertation contributes to the broader field of transfer-learning by offering a comprehensive exploration of domain generalization, domain adaptation, and close-set, and open-set settings. The insights gained from this research provide a foundation for developing effective transfer learning strategies, enabling the deployment of adaptable and high-performance models in real-world scenarios with varying domain characteristics and data availability.--Author's abstract
DOI: 10.17760/d20581914
Author: Chang Liu
Institution: Northeastern University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4388761286
