Contrastive Adaptation Network for Single- and Multi-Source Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) aims to leverage labeled source data for unlabeled target data.
• Existing UDA methods often neglect class information, leading to domain misalignment and reduced performance.
• Addressing class-specific domain discrepancies is crucial for robust UDA.

Purpose of the Study:

• Propose a novel Contrastive Adaptation Network (CAN) for UDA.
• Introduce a new metric, Contrastive Domain Discrepancy, to model intra-class and inter-class domain disparities.
• Enhance generalization performance in UDA by addressing class information.

Main Methods:

• Developed CAN to optimize Contrastive Domain Discrepancy.
• Implemented an alternating update strategy for target label estimation and feature representation.
• Utilized class-aware sampling for efficient and effective training.
• Introduced multi-source clustering ensemble and boundary-sensitive alignment for multi-source UDA.

Main Results:

• CAN demonstrated superior performance compared to state-of-the-art methods on real-world benchmarks (Office-31, VisDA-2017, DomainNet).
• Ablation studies confirmed the effectiveness of individual components within the CAN framework.
• The proposed methods achieved favorable results in both single-source and multi-source UDA scenarios.

Conclusions:

• The proposed CAN framework effectively addresses limitations of previous UDA methods by incorporating class information.
• CAN offers a robust solution for both single-source and multi-source domain adaptation.
• The novel metric and training strategies significantly improve UDA performance and generalization.