Adaptive Contrastive Learning with Label Consistency for Source Data Free Unsupervised Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) addresses domain shift but often requires source data.
• Practical constraints like privacy and intellectual property limit source data availability.
• Source-free UDA presents a more realistic challenge, adapting models without original training data.

Purpose of the Study:

• To develop a novel framework for source-free unsupervised domain adaptation.
• To enhance the discriminative power of features in the target domain.
• To propose a generalizable method applicable to various UDA tasks.

Main Methods:

• Introduced Label Consistent Contrastive Learning (LCCL), an adaptive contrastive learning framework.
• Utilized a memory bank to store pseudo-labeled and clustered target domain samples.
• Incorporated trusted historical samples into the contrastive learning process.

Main Results:

• LCCL effectively encourages target domain samples to learn class-level discriminative features.
• The proposed memory bank strategy facilitates contrastive learning without source data.
• Demonstrated the generalizability of LCCL across different UDA scenarios.

Conclusions:

• LCCL is a robust and effective approach for source-free unsupervised domain adaptation.
• The method shows significant improvements on benchmark datasets for digit recognition and image classification.
• LCCL offers a practical solution for domain adaptation challenges in privacy-sensitive applications.