Few-shot domain adaptation through compensation-guided progressive alignment and bias reduction

摘要

Most existing domain adaptation methods require large amounts of data in the target domain to train the model and a relatively long time to adapt different domains. Recently, few-shot domain adaptation (FDA) attracts lots of research attention, which only requires a small number of labeled target data and is more consistent with real-world applications. Previous works on FDA suffer from the risk of bias towards source domain and over-adapting on the target training data, which decreases the generalization of the model on the test data. In this paper, we propose a generalized framework to handle few-shot domain adaptation, named as compensation-guided progressive alignment and bias reduction (CPABR). Specifically, CPABR introduces source and target virtual data as compensations to deal with the scarcity of target data explicitly and fill in the gap between source and target domains, which promotes knowledge transfer. With the help of these virtual data, CPABR performs progressively distribution matching to gradually align the marginal and conditional distributions, and conducts weighted variance maximization to alleviate the bias of the model to the source domain. Moreover, CPABR integrates both homogeneous FDA and heterogeneous FDA into a unified framework. Extensive experiments on widely used benchmark datasets demonstrate the effectiveness of our method.