NeurIPS 2020

Domain Adaptation as a Problem of Inference on Graphical Models


Meta Review

This paper presents a novel framework for unsupervised domain adaptation. Overall reviewers and meta-reviewer appreciate the ideas of a ‘mother distribution’ but also point out that there are limitations since changes in source domains may not be reflected in target domains. A discussion on the soundness of the method under specific generative/causal model assumptions would be very useful and is strongly encouraged. Modeling conditional distributions via CGAN as proposed in Sec. 3.2 has first been proposed in CausalGAN by Kocaoglu et al. in the context of learning causal models, except for the additional theta parameter. We encourage the authors to discuss this relation.