Prototype learning for adversarial domain adaptation

Yuchun Fang

doi:10.1016/j.patcog.2024.110653

Prototype learning for adversarial domain adaptation

2024-11-01

2

PRE

AI

张兆翔 (Zhaoxiang Zhang)

S

Shaorong Xie

DOI：10.1016/j.patcog.2024.110653

Original

Original request for help

Save

Abstract

En 中文

Adversarial learning has been widely used in recent years to address the issue of domain shift in domain adaptation. However, this approach focuses on global cross -domain alignment and overlooks the alignment of class boundaries. To tackle this limitation, we introduce a novel method called PLADA. PLADA leverages prototype learning to align category distributions across domains. The prototypes in PLADA represent the source category distribution, which is constructed using labelled data and transferred to the target domain. In addition to adversarial learning for global domain -invariant feature learning, we propose the weighted prototype loss (WPL) to embed prototype information. WPL transforms the local category distribution alignment problem into a distance measurement between the prediction and prototypes, resulting in a more discriminative representation. Experimental results demonstrate that our proposed model performs comparably well on multiple classic domain adaptation tasks, showcasing the potential of PLADA.

Keywords:

Domain adaptation

Transfer learning

Unsupervised learning

Deep learning

Prototype learning for adversarial domain adaptation

Abstract

Journal

Organization