|
In real-world scenarios, machine learning tasks suffer from long-tail distribution or domain discrepancy problems, and many recent works have proposed effective methods to solve the challenges respectively. However, few studies have paid attention to the two problems simultaneously, since long-tail distribution and domain discrepancy both perhaps influence the generalization of machine learning models. Thus, according to the upper bound error theory, a design principle is given to solve the long-tail distribution with domain discrepancy problem (LT-DD) , and a pseudo-label-based decoupling domain adaptation method (PLD-DA) is proposed following the design principle in this paper.PLD-DA follows a two-stage domain adaptation framework, which trains a domain-invariant feature extractor on the original long-tail dataset at the first stage while adjusts the classifier with reweighting method at the second stage. To improve the classification confidence for the classifier, the pseudo-label information of target domain is introduced and a self-learning strategy is used. Experiments are conducted to show that our method could achieve a well-transfered feature extractor and a confident unbiased classifier simultaneously on LT-DD tasks, improves the model's generalization compared to end-end rebalancing domain adaptation methods. |
|
Keywords:Deep Learning, Domain Adaptation, Long-tail distributition, Pseudo Label |
|