How Many Domains Suffice for Domain Generalization? A Tight Characterization via the Domain Shattering Dimension
Researchers have introduced a new combinatorial measure called the domain shattering dimension to address a core question in domain generalization. This measure quantifies how many randomly sampled domains are needed to train a model that performs well across all domains within a given family. The study establishes a tight relationship between this new dimension and the classic VC dimension, proving that any hypothesis class learnable in the standard PAC framework is also learnable in this domain generalization context. AI
IMPACT Introduces a new theoretical framework for understanding model generalization across different data distributions.