Barycentric-Alignment and Reconstruction Loss Minimization for Domain Generalization

This paper advances the theory and practice of Domain Generalization (DG) in machine learning. We consider typical DG setting where hypothesis is composed a representation mapping followed by labeling function. Within this setting, majority popular methods aim to jointly learn functions minimizing well-known upper bound for classification risk unseen domain. In practice, however, based on theoretical ignore term that cannot be directly optimized due its dual dependence both unknown optimal function To bridge gap between we introduce new free terms having such dependence, resulting fully optimizable Our derivation leverages classical recent transport inequalities link metrics with information-theoretic measures. Compared previous bounds, our introduces two terms: (i) Wasserstein-2 barycenter aligns distributions domains, (ii) reconstruction loss assesses quality reconstructing original data. Based bound, propose novel algorithm named Wasserstein Barycenter Auto-Encoder (WBAE) simultaneously minimizes loss, loss. Numerical results demonstrate proposed method outperforms current state-of-the-art algorithms several datasets.