Domain Generalization: A Survey.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Machine learning models often fail on out-of-distribution (OOD) data due to the i.i.d. assumption, which is frequently violated by domain shift.
• Domain generalization (DG) addresses this by enabling models to generalize from source data to unseen target domains without access to target data during training.

Purpose of the Study:

• To provide a comprehensive literature review of domain generalization (DG) research over the past decade.
• To define DG, differentiate it from related fields like domain adaptation and transfer learning, and survey existing methodologies and theories.

Main Methods:

• Systematic review of domain generalization literature.
• Categorization of DG methods, including domain alignment, meta-learning, data augmentation, and ensemble learning.
• Analysis of DG applications across various domains such as computer vision, NLP, and medical imaging.

Main Results:

• Significant progress in DG methodologies and theoretical understanding over the last ten years.
• Identification of diverse application areas where DG has been successfully applied.
• Synthesis of the current state-of-the-art in DG research.

Conclusions:

• DG is a crucial area for achieving robust machine learning performance in real-world scenarios.
• The review highlights key advancements and identifies open challenges.
• Future research directions are proposed to further advance the field of domain generalization.