A robust quadruple adaptation network in few-shot scenarios

In few-shot domain adaptation (FDA), classifiers for target domain are trained with clean labeled data from source domain and few labeled data from target domain. However, in the source domain, it is not easy to acquire a large amount of clean labeled data in the wild world. Hence, we consider a new, more realistic and more challenging problem setting, where classifiers have to be trained with noisy labeled data from source domain and few labeled data from target domain, named wildly FDA (WFDA). We show that WFDA ruins existing FDA methods if taking no account of label noise in source domain. Therefore, we propose a robust quadruple adaptation network (QAN), a concise but effective solution to WFDA. QAN trains two models (each model consists of one encoder and one classifier) simultaneously, where one model samples data as the input of the other one to eliminate the negative impact caused by noisy source data. Experiments demonstrate that under WFDA, QAN outperforms existing baselines. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

In wildly FDA (WFDA), classifiers are trained with noisy labeled data from source domain and few labeled data from target domain. To address the issue of label noise in the source domain, a robust quadruple adaptation network (QAN) is proposed as a concise and effective solution. Experiments demonstrate that under WFDA, QAN outperforms existing baselines.