Variational Pseudo Labeling for Test Time Domain Generalization

Sameer Ambekar, Zehao Xiao, Jiayi Shen, Xiantong Zhen, Cees G. M. Snoek

Keywords: domain generalization, variational pseudo label, test-time generalization, meta-learning, probabilistic framework

TL;DR: We address test-time generalization in a probabilistic formulation by introducing variational pseudo labels

Abstract: This paper strives for domain generalization, where models are trained exclusively on source domains before being deployed at unseen target domains. We propose probabilistic pseudo labels of target samples for fine-tuning the source-trained model at test time, to generalize the model to the target domain. To do so, we formulate the adaptation as a variational inference problem by modeling pseudo labels as distributions. Variational pseudo labels are more robust to achieve a model better specified to the target domain. We learn the ability to generate better pseudo labels through simulating domain shifts during training. Experiments on widely-used datasets demonstrate the benefits, abilities and effectiveness of our proposal.

Submission Number: 10