Unsupervised multi-source domain adaptation with no observable source data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Unsupervised multi-source domain adaptation (UMDA) addresses label prediction for unlabeled target data using knowledge from multiple source domains.
• Existing UMDA methods typically require access to source data, which is often restricted due to privacy or confidentiality concerns.
• The unavailability of source data in practical scenarios necessitates novel approaches for domain adaptation.

Purpose of the Study:

• To introduce and evaluate a novel approach for data-free unsupervised multi-source domain adaptation (UMDA).
• To develop a method that enables knowledge transfer from multiple source domains to a target domain without accessing any source data.
• To address the realistic challenge of UMDA when source data is inaccessible.

Main Methods:

• Propose DEMS (Data-free Exploitation of Multiple Sources), a novel architecture for data-free UMDA.
• Adapt target data to source domains without utilizing any source data.
• Estimate target labels by leveraging pre-trained source classifiers.

Main Results:

• DEMS achieves state-of-the-art accuracy in data-free UMDA tasks.
• The proposed method demonstrates significant performance improvements, with accuracy gains up to 27.5% compared to existing baselines.
• Extensive experiments on real-world datasets validate the effectiveness of DEMS.

Conclusions:

• DEMS effectively solves the data-free UMDA problem, a novel and crucial area.
• The architecture successfully adapts target data and predicts labels without source data access.
• DEMS represents a significant advancement in domain adaptation under data privacy constraints.