Understanding Self-Training for Gradual Domain Adaptation

2022년 04월 01일 (금) | 발표자: 임진혁

This paper emphasizes the necessity of the gradual shift problem in which domain shift occurs gradually over time among unsupervised domain adaptation. Through theoretical analysis, the authors analyze when the gradual structure proposed by the paper is helpful for target domain error and present a reliable algorithm based on this. Higher accuracy is obtained than a direct unsupervised adaptation for various datasets.