Domain knowledge boosted adaptation: Leveraging vision-language models for multi-source domain adaptation

Yuwei He

doi:10.1016/j.neucom.2024.129114

Domain knowledge boosted adaptation: Leveraging vision-language models for multi-source domain adaptation

2025-02-01

0

PRE

AI

丁贵广 (Guiguang Ding) *

DOI：10.1016/j.neucom.2024.129114

Original

Original request for help

Save

Abstract

En 中文

Multi-source domain adaptation (MSDA) aims to adapt a model trained on multiple labeled source domains to an unlabeled target domain. Existing MSDA methods primarily focus on reducing domain gaps by aligning the source domains with the target domain, either jointly or separately. However, these methods often distort semantic-related features and overlook the valuable domain-related information present in diverse domains. In this paper, we propose a novel MSDA method called Domain Knowledge Boosted Adaptation (DKBA) that leverages domain-related information to enhance model performance. Firstly, we employ prompt learning to embed domain-related information learned from a pretrained vision-language model into prompt embeddings. These embeddings serve as conditional priors, allowing the classification model to adaptively embed semantic- related features and obtain domain-invariant semantic features without excessively aligning domains. Our proposed DKBA approach achieves state-of-the-art results on four MSDA datasets, highlighting its effectiveness in leveraging domain knowledge for improved adaptation performance.

Keywords:

Deep learning

Multi-source domain adaptation

Prompt learning

Domain knowledge boosted adaptation: Leveraging vision-language models for multi-source domain adaptation

Abstract

Journal

Organization