Addressing Data Scarcity in Domain Generalization for Computer Vision Applications in Image Classification

Abstract

Domain generalization (DG) for image classification is a crucial task in machine learning that focuses on transferring domain-invariant knowledge from multiple source domains to an unseen target domain. Traditional DG methods assume that classes of interest are present across multiple domains (domain-shared), which helps mitigate spurious correlations between domain and class. However, in real-world scenarios, data scarcity often leads to classes being present in only a single domain (domain-linked), resulting in poor generalization performance. This thesis introduces the domain-linked DG task and proposes a novel methodology to address this challenge.

This thesis proposes FOND, a "Fairness-inspired cONtrastive learning objective for Domain-linked domain generalization," which leverages domain-shared classes to learn domain-invariant representations for domain-linked classes. FOND is designed to enhance generalization by minimizing the impact of task-irrelevant domain-specific features.

The theoretical analysis in this thesis extends existing domain adaptation error bounds to the domain-linked DG task, providing insights into the factors that influence generalization performance. Key theoretical findings include the understanding that domain-shared classes typically have more samples and learn domain-invariant features more effectively than domain-linked classes. This analysis informs the design of FOND, ensuring that it addresses the unique challenges of domain-linked DG.

Furthermore, experiments are performed across multiple datasets and experimental settings to evaluate the effectiveness of various current methodologies. The proposed method achieves state-of-the-art performance in domain-linked DG tasks, with minimal trade-offs in the performance of domain-shared classes. Experimental results highlight the impact of shared-class settings, total class size, and inter-domain variations on the generalizability of domain-linked classes. Visualizations of learned representations further illustrate the robustness of FOND in capturing domain-invariant features.

In summary, this thesis advocates future DG research for domain-linked classes by (1) theoretically and experimentally analyzing the factors impacting domain-linked class representation learning, (2) demonstrating the ineffectiveness of current state-of-the-art DG approaches, and (3) proposing an algorithm to learn generalizable representations for domain-linked classes by transferring useful representations from domain-shared ones.