A1238
Title: Stein discrepancy for unsupervised domain adaptation
Authors: Anneke von Seeger - University of Minnesota - Twin Cities (United States) [presenting]
Gilad Lerman - University of Minnesota - Twin Cities (United States)
Dongmian Zou - Duke Kunshan University (China)
Abstract: Unsupervised domain adaptation (UDA) aims to improve model performance on an unlabeled target domain using a related, labeled source domain. A common approach aligns source and target feature distributions by minimizing a distance between them, often using symmetric measures such as maximum mean discrepancy (MMD). However, these methods struggle when target data is scarce. A novel UDA framework is proposed that leverages Stein discrepancy, an asymmetric measure that depends on the target distribution only through its score function, making it particularly suitable for low-data target regimes. The proposed method has kernelized and adversarial forms and supports flexible modeling of the target distribution via Gaussian, GMM, or VAE models. A generalization bound on the target error and a convergence rate for the empirical Stein discrepancy in the two-sample setting are derived. Empirically, the method consistently outperforms prior UDA approaches under limited target data across multiple benchmarks.
