2019/03/07 by Subhankar Roy, Aliaksandr Siarohin, Roy, Subhankar +9 · 7 citations
Computer Science · Medicine · #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1903.03215
openalex publication_date 2019/03/07 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
A classifier trained on a dataset seldom works on other datasets obtained\nunder different conditions due to domain shift. This problem is commonly\naddressed by domain adaptation methods. In this work we introduce a novel deep\nlearning framework which unifies different paradigms in unsupervised domain\nadaptation. Specifically, we propose domain alignment layers which implement\nfeature whitening for the purpose of matching source and target feature\ndistributions. Additionally, we leverage the unlabeled target data by proposing\nthe Min-Entropy Consensus loss, which regularizes training while avoiding the\nadoption of many user-defined hyper-parameters. We report results on publicly\navailable datasets, considering both digit classification and object\nrecognition tasks. We show that, in most of our experiments, our approach\nimproves upon previous methods, setting new state-of-the-art performances.\n