Hierarchical Domain Invariant Variational Auto-Encoding with weak domain supervision

Abstract

We address the task of domain generalization, where the goalis to train a predictive model such that it is able to generalizeto a new, previously unseen domain. We choose a generativeapproach within the framework of variational autoencodersand propose an unsupervised algorithm that is able to gen-eralize to new domains without supervision. We show thatour method is able to learn representations that disentangledomain-specific information from class-label specific infor-mation even in complex settings where domain structure isnot observed during training. Our interpretable method out-performs previously proposed generative algorithms for do-main generalization and achieves competitive performancecompared to state-of-the-art approaches, which rely on ob-serving domain-specific information during training, on thestandard domain generalization benchmark dataset PACS.Additionally, we proposed weak domain supervision whichcan further increase the performance of our algorithm in thePACS dataset.

Publication
ICLR 2021 Workshop on robustML
Florian Buettner
Florian Buettner
Professor for Bioinformatics in Oncology